Next Article in Journal
Calculation of Greenhouse Gas Emissions from Tourist Vehicles Using Mathematical Methods: A Case Study in Altai Tavan Bogd National Park
Previous Article in Journal
Study on Impact of Managerial Effectiveness and Digitalization on Green Total Factor Productivity of Enterprises: Sample of Listed Heavy-Polluting Enterprises in China
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Spatio-Temporal Variation in Solar Irradiance in the Mediterranean Region: A Deep Learning Approach

1
Department of Software Engineering, Istanbul Topkapi University, 34087 Istanbul, Turkey
2
Department of Computer Information Systems and Business Analytics, Metropolitan State University of Denver, Denver, CO 80217, USA
3
Faculty of Science and Literature, Süleyman Demirel University, 32260 Isparta, Turkey
4
Department of Computer Engineering, Faculty of Engineering, Istanbul Aydın University, 34295 Istanbul, Turkey
5
VNC, International Staff—IHC, 1110 Brussels, Belgium
*
Author to whom correspondence should be addressed.
Sustainability 2025, 17(15), 6696; https://doi.org/10.3390/su17156696
Submission received: 1 June 2025 / Revised: 18 July 2025 / Accepted: 19 July 2025 / Published: 23 July 2025
(This article belongs to the Section Energy Sustainability)

Abstract

In response to the global imperative of reducing greenhouse gas emissions, the optimisation of renewable energy systems under regionally favourable conditions has become increasingly essential. Solar irradiance forecasting plays a pivotal role in enhancing energy planning, grid reliability, and long-term sustainability. However, in the context of Turkey, existing studies on solar radiation forecasting often rely on traditional statistical approaches and are limited to single-site analyses, with insufficient attention to regional diversity and deep learning-based modelling. To address this gap, the present study focuses on Turkey’s Mediterranean region, characterised by high solar potential and diverse climatic conditions and strategically relevant to national clean energy targets. Historical data from 2020 to 2023 were used to forecast solar irradiance patterns up to 2026. Five representative locations—Adana, Isparta, Fethiye, Ulukışla, and Yüreğir—were selected to capture spatial and temporal variability across inland, coastal, and high-altitude zones. Advanced deep learning models, including artificial neural networks (ANN), long short-term memory (LSTM), and bidirectional LSTM (BiLSTM), were developed and evaluated using standard performance metrics. Among these, BiLSTM achieved the highest accuracy, with a correlation coefficient of R = 0.95, RMSE = 0.22, and MAPE = 5.4% in Fethiye, followed by strong performance in Yüreğir (R = 0.90, RMSE = 0.12, MAPE = 7.2%). These results demonstrate BiLSTM’s superior capacity to model temporal dependencies and regional variability in solar radiation. The findings contribute to the development of location-specific forecasting frameworks and offer valuable insights for renewable energy planning and grid integration in solar-rich environments.

1. Introduction

Solar irradiance forecasting plays a crucial role in advancing renewable energy systems by optimising solar power technologies and supporting the strategic planning of sustainable energy infrastructure. Accurate predictions are essential for enhancing energy conversion efficiency, facilitating effective energy storage, ensuring grid stability, and improving load balancing, thereby contributing to the transition towards cleaner energy systems. However, the inherent variability of solar irradiance, influenced by atmospheric phenomena, geographical heterogeneity, and seasonal cycles, presents significant challenges. These complexities necessitate the development of advanced predictive models capable of capturing intricate temporal and spatial dependencies [1,2].
Given these challenges, countries with high solar potential, such as Turkey, require accurate forecasting models to maximise energy yield and ensure efficient integration into the power grid. Turkey possesses substantial solar energy potential due to its favourable geographical location, with an average annual solar radiation of approximately 1527 kWh/m2 and 2737 h of sunshine per year. Recognising this potential, the country has set ambitious targets to expand its solar energy capacity. According to the National Energy Plan, Turkey aims to increase its installed solar photovoltaic (PV) capacity to 30 GW by 2035, with a long-term goal of 300 GW by 2050 [3]. This substantial expansion is expected to contribute significantly to the decarbonisation of the energy sector, reducing dependence on fossil fuels and enhancing energy security. In addition to large-scale solar farms, distributed PV applications, such as rooftop installations, solar-integrated buildings, and floating solar panels, are being encouraged to support decentralised electricity generation. However, achieving these targets requires accurate solar forecasting models to mitigate the challenges posed by fluctuations in solar irradiance, ensuring reliable grid integration and optimal energy management.
The Weather Research and Forecasting (WRF) model has been widely employed in solar irradiance forecasting, particularly for evaluating global horizontal irradiance (GHI). However, its accuracy is highly sensitive to atmospheric conditions and parameterisation schemes. These limitations have motivated the exploration of alternative methods, such as statistical models and machine learning (ML) techniques, which offer improved adaptability and performance in capturing nonlinear irradiance patterns [4,5,6,7,8,9].
More recently, deep learning (DL) methods have emerged as an even more powerful alternative, as they can extract complex temporal features directly from data without the need for domain-specific parametrisation.
The advent of DL models has revolutionised solar irradiance forecasting by introducing sophisticated tools capable of capturing intricate temporal dependencies in time-series data. Artificial neural networks (ANNs) have significantly improved forecasting by effectively modelling complex nonlinear relationships within solar irradiance data. Their adaptability and ability to handle multi-variable datasets have led to superior predictive accuracy compared to traditional statistical methods [10,11,12,13,14,15,16]. Recent studies, such as Chodakowska (2024), have underscored the substantial contributions of ANNs to solar irradiance forecasting. These studies highlight the adaptability of ANNs in modelling nonlinear relationships and their superior predictive accuracy over conventional statistical approaches [17]. Furthermore, applications of ANN-based forecasting have been extended to predict solar photovoltaic (PV) power output and estimate battery state of charge, demonstrating the flexibility of ANN models in renewable energy management [18,19]. However, despite these advancements, ANNs struggle to capture long-term temporal dependencies, which diminishes their effectiveness in sequential tasks requiring extended time-series analysis. This limitation has prompted the adoption of advanced architectures, such as long short-term memory (LSTM) and bidirectional LSTM (BiLSTM) networks, which offer enhanced capabilities for modelling temporal dynamics in solar irradiance data.
Building on the limitations of ANNs in handling long-term temporal dependencies, the emergence of LSTM networks has addressed these challenges by introducing architectures specifically designed for sequential data [20,21,22,23,24,25,26]. LSTM networks were specifically designed to address the vanishing gradient problem frequently encountered in traditional recurrent neural networks, thereby enabling the retention of relevant temporal information over extended sequences [27,28]. This architecture has proven particularly effective for time-series forecasting tasks, including solar irradiance prediction, where both short- and long-term dependencies must be captured with precision. Recent studies have empirically demonstrated the superiority of LSTM models over conventional statistical methods and feedforward neural networks, especially in dynamic environments requiring robust temporal pattern recognition [29,30]. For instance, Campos et al. [31] aimed to forecast short-term photovoltaic (PV) energy production using LSTM architectures. In their study, real-time solar irradiance and temperature data were collected from a PV installation in Brazil, and an LSTM model was trained to perform energy output forecasting with a 24 h lead time. The results demonstrated that the LSTM model significantly outperformed feedforward ANN and linear regression benchmarks, achieving a mean absolute error (MAE) reduction of up to 22%, thereby confirming its ability to model nonlinear and sequential patterns in solar data more effectively. Similarly, Ehteram et al. [32] introduced a novel LSTM variant called the “Read-First LSTM” (RF-LSTM), designed specifically to enhance long-term feature retention in solar radiation datasets. The study utilised historical solar radiation data from three distinct climatic zones in Iran and compared the predictive performance of RF-LSTM against standard LSTM and gated recurrent unit (GRU) models. The proposed RF-LSTM achieved the lowest root mean square error (RMSE) across all regions, demonstrating superior robustness and generalisability for varying temporal patterns. The authors concluded that enhanced LSTM configurations can significantly reduce forecasting errors and increase the reliability of solar energy management systems. These advancements have positioned LSTM as a pivotal tool in renewable energy research, addressing the limitations of previous approaches and paving the way for more reliable forecasting models [33,34].
While LSTM models process input sequences in a unidirectional manner, bidirectional LSTM (BiLSTM) networks improve upon this architecture by capturing both past and future dependencies, offering a more comprehensive temporal modelling framework for solar irradiance forecasting. This structure enables the model to learn intricate temporal relationships more effectively, especially in data influenced by fluctuating weather conditions. Recent studies have demonstrated the effectiveness of BiLSTM networks in capturing both short-term and long-term dependencies. For example, Peng et al. [35] integrated BiLSTM with the sine cosine algorithm (SCA) for hyperparameter tuning, achieving an RMSE of 61.38 W/m2 and MAE of 44.07 W/m2, outperforming conventional LSTM and ANN models by 15–20% [35]. Similarly, Bou-Rabee et al. [36] evaluated BiLSTM under clear and cloudy conditions in continental climate zones, reporting an MAE of 35.72 W/m2 and RMSE of 48.15 W/m2, surpassing baseline RNNs and GRUs. Further, Li et al. [37] developed a BiLSTM model incorporating daily irradiance amplitude as input, achieving an RMSE of 51.6 W/m2 and MAE of 38.9 W/m2 across three climatic regions in China. Additional works demonstrated BiLSTM’s capacity in day-ahead irradiance prediction and even in operational contexts such as solar panel maintenance scheduling [38,39]. These empirical findings highlight BiLSTM’s versatility and its superior ability to adapt to dynamic irradiance profiles. A summary of these approaches, including input parameters, model types, and key error metrics (MAE, RMSE), is provided in Table 1.
Turkey’s geographical location offers considerable solar energy potential; however, its utilisation remains relatively limited, with current applications predominantly concentrated in domestic water heating. The adoption of innovative and data-driven approaches to solar energy forecasting has yet to reach widespread implementation. Bridging this gap through accurate and region-specific solar irradiance forecasting is essential for optimising solar energy systems, enhancing grid integration, and supporting sustainable energy policies. This study presents a comparative analysis of advanced deep learning models—artificial neural networks (ANN), long short-term memory (LSTM), and bidirectional LSTM (BiLSTM)—to forecast the spatio-temporal variation in solar irradiance across Turkey’s Mediterranean region. While previous studies have typically focused on single-site analyses or the use of a single predictive model, such approaches often lack the capacity to capture regional climatic diversity or to benchmark the relative strengths of different modelling frameworks. In contrast, the present research integrates multiple models across five geographically and climatically distinct locations—Adana, Isparta, Fethiye, Ulukışla, and Yüreğir—using historical irradiance data from 2020 to 2023 to forecast trends through 2026. The selection of these locations enables the modelling framework to reflect variations in elevation, topography, and microclimatic conditions, thereby increasing the generalisability and robustness of the predictions. The comparative evaluation highlights not only the predictive performance of each model but also their specific advantages and limitations, with particular emphasis on the enhanced temporal learning capacity of BiLSTM through bidirectional sequence processing.
This study offers several original contributions: (i) it presents the first multi-regional, deep learning-based spatio-temporal forecasting of solar irradiance in the Mediterranean region of Turkey; (ii) it systematically compares three advanced deep learning techniques, offering a detailed performance analysis that informs model selection for future studies; and (iii) it provides a practical foundation for policymakers and energy planners by enhancing forecasting accuracy, thereby contributing to the optimisation of solar energy utilisation and grid integration. The findings of this research advance the current understanding of solar irradiance prediction and support the development of more efficient and sustainable renewable energy strategies.
To summarise, the main problem addressed in this study is the insufficient accuracy and generalisability of solar irradiance forecasting models in regions with complex geographical and climatic variability, such as Turkey’s Mediterranean region. The primary objective is to enhance the spatio-temporal precision of irradiance forecasts through a comparative analysis of three deep learning techniques: ANN, LSTM, and BiLSTM. To achieve this, a series of tasks were undertaken, including the collection of historical irradiance data (2020–2023) from five diverse locations, the development and training of each model, and the assessment of their predictive performance using standard evaluation metrics. The novelty of this study lies in its multi-regional, deep learning-based forecasting framework and its systematic comparison of sequential and bidirectional architectures. The practical significance is reflected in the model’s potential to inform photovoltaic system planning, support regional grid integration, and contribute to national solar energy targets by providing location-specific, data-driven insights.

2. Materials and Methods

2.1. Study Area

This research focuses on analysing the extensive impacts of climate change on solar irradiance and solar energy potential in Turkey’s Mediterranean region. The selected regions, from both the western and eastern parts of the Mediterranean, are representative measured points from the Mediterranean region of Turkey. The map of the study regions, illustrating the areas selected from both the western and eastern parts of the Mediterranean, is presented in Figure 1 and was created using ArcGIS Pro version 3.2 (Esri, Redlands, CA, USA).
Yüreğir, located in southern Adana near the Taurus Mountains, exhibits a transitional Mediterranean climate. Its proximity to coastal plains and river systems contributes to high solar exposure, although occasional cloud cover and humidity influence irradiance variability.
Ulukışla, in Niğde province, lies at the intersection of the Mediterranean and Central Anatolian climatic zones, at an elevation of 1427 m. Its high altitude and inland position result in significant daily and seasonal temperature variation, affecting solar radiation patterns.
Fethiye, in western Muğla province, represents a coastal Mediterranean setting with mild winters and hot, dry summers. The region benefits from high sunshine duration and relatively low cloud cover, making it favourable for solar energy applications.
Adana (37°00′ N, 35°19′ E), located in the Çukurova plain, has a typical Mediterranean climate characterised by hot summers (mean daily maximum ~31 °C), high nocturnal humidity (>85%), and limited summer rainfall. With ~87.5% sunshine during daylight hours annually, Adana offers high solar energy potential [40].
Isparta, in the northern Mediterranean zone, is situated at an average elevation of 1050 m and exhibits a mixed climate, influenced by both Mediterranean and continental conditions. It features cold winters (January avg. 1.8 °C), warm summers (July avg. 23.5 °C), and moderate annual precipitation (~600 mm), mostly occurring between December and May.

2.2. Data Used

The solar irradiance data used in this study were collected between 2020 and 2023 from five regions in Turkey’s Mediterranean region. These data were obtained from the Turkish State Meteorological Service (TSMS), ensuring their reliability and accuracy for analysis. The dataset reveals notable seasonal and spatial variations, with the highest solar irradiance levels observed during summer. July demonstrated peak values across all regions, reaching 7.75 kWh/m2/day in Fethiye and Isparta. Conversely, the lowest irradiance levels occurred during winter, particularly in December, with Isparta recording the minimum at 1.92 kWh/m2/day. Among the regions studied, Fethiye consistently exhibited the highest solar irradiance levels during the summer, while Isparta showed significantly lower values during winter. These findings highlight the variability in solar irradiance across the Mediterranean region, providing critical insights into its solar energy potential (Table 2).
The annual solar irradiance levels between 2020 and 2023 across Yüreğir, Ulukışla, Fethiye, Adana, and Isparta reveal consistent patterns with slight year-to-year fluctuations (Table 3). Yüreğir experienced its highest value in 2021 at 5.09 kWh/m2/day, followed by a gradual decline to 4.85 kWh/m2/day in 2023. Similarly, Ulukışla peaked in 2021 at 4.81 kWh/m2/day and decreased to 4.61 kWh/m2/day by 2023. Fethiye demonstrated stable levels, with a slight peak in 2021 at 5.07 kWh/m2/day, decreasing marginally to 4.94 kWh/m2/day in 2023. Adana showed minimal variation, maintaining consistently low levels between 4.50 and 4.59 kWh/m2/day. Isparta, which recorded the highest average in 2020 at 5.14 kWh/m2/day, saw a steady decrease, reaching 4.67 kWh/m2/day in 2023. These variations suggest stable solar irradiance potential in the region, with 2021 being a slightly more favourable year for most regions, likely due to specific atmospheric or climatic conditions. However, the gradual decline in some regions, particularly Isparta, may indicate minor shifts in local weather patterns or data variability. Overall, the results demonstrate a reliable solar energy potential across all regions, with year-to-year changes being relatively insignificant.

2.3. Methodology

Figure 2 illustrates the preprocessing and forecasting pipeline described in this section. This study aimed to forecast daily solar irradiance values up to 31 December 2026 by employing deep learning techniques on historical observations. A dataset comprising 1393 daily irradiance measurements, recorded by the Turkish State Meteorological Service (TSMS) between 1 January 2020 and 23 October 2023, was utilised. The raw dataset was initially inspected for anomalies. Outliers were identified using the interquartile range (IQR) criterion, where values beyond 1.5 × IQR from the lower and upper quartiles were removed.
Missing data points, which constituted less than 2% of the dataset, were rare and were addressed using linear interpolation based on adjacent temporal values, thereby ensuring temporal continuity without distorting trend information. To reduce high-frequency noise, a three-day rolling mean filter was applied. Subsequently, the data were standardised using Z-score normalisation to centre the values around zero and scale them to unit variance. This transformation ensured that approximately 99.7% of the normalised values fell within three standard deviations of the mean, thereby enhancing model stability and convergence. A direct one-step forecasting strategy was adopted. For each day t, the input xt was defined as the observed irradiance on day t, and the target yt+1 was defined as the irradiance on day t + 1. This approach yielded 1392 supervised training samples. For multi-day forecasting between 24 October 2023 and 31 December 2026, an iterative prediction method was employed. In this approach, each predicted value was appended to the input series and used to predict the next value, thereby simulating forward projection. The dataset was chronologically partitioned to simulate operational forecasting conditions. Observations from 1 January 2020 to 31 August 2023 were used for training, and data from 1 September 2023 to 23 October 2023 were reserved for testing and performance evaluation. To further validate the robustness and temporal generalisability of the models, a five-fold time-series cross-validation was conducted. In each fold, the earliest 80% of the training data were used for model training, and the subsequent 20% of the training data were used for validation. Although higher-frequency irradiance data (e.g., hourly) might offer finer temporal resolution and potentially enhanced model precision, only daily averaged values were consistently available across all five study regions for the observation period. As a result, the modelling framework adopted daily resolution to ensure uniformity and comparability across locations.

2.3.1. ANN

ANNs are computational models inspired by the structure and functionality of biological neural networks. These models consist of interconnected nodes, or neurons, organised in layers, which process input data through weighted connections and activation functions to produce outputs. The general structure of an ANN can be represented mathematically as in Equation (1):
y = f 1 n w i x i   +   b
where y represents the output, xi are the input variables, wi are the weights, b is the bias term, and f denotes the activation function. The learning process involves adjusting xi and b through backpropagation to minimise the error between predicted and actual outputs.
In this study, ANNs were utilised due to their capability to model the nonlinear and complex interactions present in solar irradiance data. Traditional statistical methods often struggle with such data due to its inherent variability and dynamic nature. ANNs, on the other hand, adaptively learn from data, enabling high prediction accuracy and robustness [41]. Additionally, their ability to handle large datasets and capture intricate relationships between meteorological variables makes them ideal for forecasting solar irradiance [42]. By leveraging these properties, ANN-based models have been shown to reduce error rates and enhance prediction reliability in renewable energy forecasting applications [43].

2.3.2. LSTM

LSTM networks are a specialised recurrent neural network (RNN) designed to capture both short-term and long-term dependencies in sequential data effectively. Unlike traditional RNNs, LSTMs utilise a gating mechanism to regulate information flow, thereby mitigating issues such as the vanishing gradient problem and enabling the retention of pertinent information over extended time intervals. This architecture renders LSTMs particularly adept at modelling temporal sequences where context from earlier data points is crucial for accurate predictions.
This study employed LSTM networks to forecast solar irradiance due to their proficiency in handling the temporal and nonlinear characteristics inherent in meteorological data. Recent research has demonstrated that LSTMs outperform traditional machine learning models in predicting solar irradiance, owing to their capacity to model complex temporal patterns [44]. Furthermore, LSTMs have shown superior performance in short-term solar irradiance forecasting compared to other time-series methods, attributed to their ability to capture intricate dependencies within the data [45]. The robustness of LSTMs in managing missing or noisy data, common in meteorological datasets, further enhances their suitability for this application [46]. Additionally, the adaptability of LSTMs to incorporate multiple input features, such as temperature and humidity, allows for a comprehensive analysis of factors influencing solar irradiance, leading to more accurate and reliable forecasts [47].

2.3.3. BiLSTM

BiLSTM networks are an advanced extension of traditional LSTM architectures, designed to process sequential data in both forward and backward directions. This bidirectional approach allows for BiLSTMs to capture contextual information from both past and future states, enhancing their capacity to model complex temporal dependencies within data sequences. By integrating two LSTM layers—one processing the sequence from start to end and the other from end to start—BiLSTMs utilise information from the entire sequence, making them particularly effective for tasks where surrounding context is critical [48].
In this study, BiLSTM networks were employed to forecast solar irradiance, as they offer significant advantages in capturing intricate temporal patterns in meteorological data. Unlike traditional methods, which may fail to utilise the sequential dependencies in time-series data fully, BiLSTMs excel at leveraging bidirectional information, providing a more comprehensive understanding of the temporal dynamics. Recent studies have demonstrated that BiLSTM models outperform unidirectional LSTMs in predicting solar irradiance due to their ability to account for preceding and succeeding time steps, thereby improving forecasting accuracy [49]. Additionally, their ability to model bidirectional dependencies enables BiLSTMs to achieve enhanced performance in short-term solar energy forecasting compared to traditional machine learning methods [50].
The choice of BiLSTM for this study is further supported by its robustness in handling complex sequential patterns, which are often present in solar irradiance data influenced by varying atmospheric conditions. This capability ensures stable and reliable predictions, even in challenging datasets. Moreover, the adaptability of BiLSTMs to integrate multiple input features, such as temperature and humidity, makes them a valuable tool for comprehensive solar irradiance modelling. These strengths make BiLSTMs highly effective in renewable energy applications, facilitating accurate and reliable solar irradiance predictions, which are critical for optimising solar energy systems [51].

2.3.4. Model Fit Measures

In this study, predictive modelling was conducted using neural network architectures, including ANN, LSTM, and BiLSTM, and the models’ performance was evaluated using multiple error metrics. MSE and RMSE were employed to quantify the average magnitude of prediction errors, with RMSE being particularly useful as it expresses errors in the same units as the observed values, thereby facilitating direct comparisons of model performance. MAPE was used to evaluate forecast accuracy in relative terms, offering insights into the proportional magnitude of errors with respect to the actual values. In addition, the R2 was calculated to assess the strength of the relationship between the predicted and actual values, thus providing a clear indication of the models’ reliability in capturing temporal patterns [52,53]. These metrics were selected to ensure a comprehensive evaluation of the models, addressing both absolute and relative errors while highlighting the degree of alignment between predictions and observed outcomes. Their mathematical formulations are presented below:
M S E   =   1 n ×   i = 1 n y i     ŷ i 2
R M S E = 1 n × i = 1 n y i ŷ i 2    
M A P E = 100 % n i = 1 n y i ŷ i y i
R 2 = 1 i = 1 n y i ŷ i 2 i = 1 n y i ȳ 2  
Here, yᵢ denotes the actual value, ŷᵢ denotes the predicted value, ȳ denotes the mean of actual values, and n denotes the total number of observations.

3. Results

The dataset employed in this study was partitioned into three temporally consistent subsets: training (1 January 2020–30 August 2023), testing (1 September 2023–23 October 2023), and validation (24 October 2023–31 December 2023). Subsequent forecasts were extended through to the end of 2026. This temporal segmentation facilitated a robust assessment of model performance across distinct periods, thereby ensuring the reliability of the predictions. The analysis delineates solar irradiance trends across multiple locations in the Mediterranean region, capturing the area’s varied climatic characteristics.

3.1. Outputs of ANN Model

The ANN implemented in this study is based on a nonlinear autoregressive network with exogenous inputs (NARX). The architecture comprises an input layer representing daily solar irradiance values, followed by a single hidden layer. The number of neurons in the hidden layer was empirically determined through iterative testing to balance the risks of underfitting and overfitting. The ReLU activation function was employed in the hidden layer to introduce nonlinearity. The network was trained using the Levenberg–Marquardt optimisation algorithm, which is particularly effective for small- to medium-sized datasets. Mean squared error (MSE) was used as the loss function. The training process was carried out for 100 epochs with a batch size of 32.
Although no automated optimisation algorithm was used, hyperparameters were selected based on experimental tuning via a trial-and-error strategy informed by validation performance. This approach was considered appropriate given the dataset size and computational constraints.
Model performance was evaluated with standard metrics, including MAPE, MSE, RMSE, and R. The ANN model achieved the highest accuracy in Fethiye, with a correlation coefficient of 0.73 and lowest error values (MSE = 0.3, RMSE = 0.61, MAPE = 8.2%), indicating strong capability in modelling consistent solar irradiance patterns. Moderate performance was observed in Isparta (R = 0.69, MAPE = 17.6%) and Yüreğir (R = 0.65, MAPE = 16.8%), while predictive accuracy declined in Adana (R = 0.64, MAPE = 24%) and Ulukışla (R = 0.45, MAPE = 23.3%), likely due to increased climatic variability. These results highlight the model’s ability to capture general temporal trends effectively, though additional input features or architectural refinements may be necessary to improve forecasting accuracy in regions with higher variability (Table 4).
The following figures (Figure 3, Figure 4, Figure 5, Figure 6 and Figure 7) present the actual, test, and forecasted annual average solar irradiance values generated by the ANN model for 2020 to 2026. The blue section represents the actual historical data from 2020 to 2023, while the red section shows the model’s test phase performance, where predictions were validated against actual data from the same period. Finally, the green section illustrates the forecasted values for 2023 to 2026 as predicted by the ANN model. These visualisations provide a comprehensive assessment of the model’s ability to identify temporal patterns and predict future trends in solar irradiance for the specified regions. Based on the ANN analysis, the forecasted annual average solar irradiance levels for 2026 are as follows: Yüreğir is predicted to reach 4.20 kWh/m2/day, with an error rate of 16.8%; Ulukışla is forecasted at 4.34 kWh/m2/day, with an error rate of 23.3%; Fethiye is expected to achieve 4.70 kWh/m2/day, with the lowest error rate of 8.2%; Isparta is predicted at 3.83 kWh/m2/day, with an error rate of 17.6%; and Adana is forecasted to record 4.30 kWh/m2/day, with the highest error rate of 24%. These results demonstrate the ANN model’s ability to estimate annual solar irradiance levels while accounting for regional differences and associated error margins. The comparison between actual and predicted values provides valuable insights into the model’s predictive accuracy and highlights areas for further optimisation.

3.2. Outputs of LSTM Model

The LSTM model architecture comprised three stacked LSTM layers, each with 128 hidden units. To mitigate overfitting, a dropout layer with a rate of 0.2 was applied after each LSTM layer. The output from the final LSTM layer was passed to a fully connected dense layer with a linear activation function to produce the forecasted solar irradiance values. The model was trained using the Adam optimiser with a learning rate of 0.001, a batch size of 32, and for 100 epochs. These hyperparameters—including the number of hidden units, dropout rate, learning rate, and training configuration—were selected through manual iterative tuning. Specifically, combinations of key parameters were tested sequentially to observe their influence on validation loss, and the configuration yielding the best generalisation performance on the validation fold was adopted. Grid search or automated methods were not employed due to computational constraints and the relatively limited dataset size. This empirical tuning approach ensured that hyperparameters were not arbitrarily chosen, thus supporting model reproducibility within the context of the dataset used. Given the daily resolution and the independent nature of samples across batches, a stateless LSTM configuration was adopted. As the forecasting strategy did not require the preservation of hidden states across batches, stateful LSTM was not considered.
Table 5, which summarises the performance metrics of the LSTM model for solar irradiance forecasting, reveals variations in the model’s predictive accuracy across regions. The model performed best in Fethiye, achieving the highest correlation (R = 0.83) and the lowest error metrics (MSE = 0.56, RMSE = 0.39, MAPE = 7.9%). This demonstrates the model’s ability to capture the consistent solar irradiance patterns in the region effectively. In Yüreğir, the model showed moderate accuracy, with R = 0.68, MSE = 0.52, RMSE = 0.72, and MAPE = 15.1%. Similarly, Isparta displayed strong predictive capabilities, with R = 0.74, MSE = 0.78, RMSE = 0.62, and MAPE = 15.4%, indicating acceptable performance. However, the model faced challenges in Ulukışla, where R = 0.61, the lowest among all regions, and error metrics were higher (MSE = 1.08, RMSE = 1.17, MAPE = 21.03%). Adana also presented difficulties despite having R = 0.74. The error values (MSE = 0.82, RMSE = 0.68, MAPE = 22.3%) suggest that the model struggled to accurately capture the region’s temporal solar irradiance patterns.
The following figures (Figure 8, Figure 9, Figure 10, Figure 11 and Figure 12) display the actual, test, and forecasted solar irradiance values obtained using the LSTM model between 2020 and 2026. The blue section represents the actual historical solar irradiance data from 2020 to 2023. The red section shows the model’s performance during the test phase, where predictions are validated against actual data from the same period. Finally, the green section illustrates the forecasted solar irradiance values from 2023 to 2026 as predicted by the LSTM model. These figures help evaluate the model’s ability to capture temporal patterns in solar irradiance and forecast future trends in the specified region. The comparison between the actual and predicted values provides valuable insights into the accuracy of the LSTM model. Based on the LSTM analysis, the forecasted annual average solar irradiance levels for 2026 are as follows: Yüreğir is predicted to have 4.90 kWh/m2/day, with an error rate of 15.1%; Ulukışla is forecasted at 4.72 kWh/m2/day, with an error rate of 21.4%; Fethiye is expected to reach 5.3 kWh/m2/day, with an error rate of 7.9%; Isparta is predicted to record 5.80 kWh/m2/day, with an error rate of 15.4%; and Adana is forecasted at 4.70 kWh/m2/day, with an error rate of 22.3%. These values highlight the model’s capability in estimating annual solar irradiance averages while accounting for regional variations and associated error rates.

3.3. BİLSTM Model

The BiLSTM model was trained on historical daily solar irradiance data using a multi-layer architecture optimised for time-series forecasting. It comprised two consecutive bidirectional LSTM layers, each with 50 hidden units. To prevent overfitting, a dropout layer with a rate of 0.5 was applied after the BiLSTM layers. The output was then passed to a fully connected dense layer with a linear activation function to generate the forecasted irradiance values. The model was trained using the Adam optimiser with an initial learning rate of 0.0001, a mini-batch size of 64, and for 200 epochs. Mean squared error (MSE) was employed as the loss function.
Hyperparameters—including the number of hidden units, dropout rate, learning rate, and batch size—were not arbitrarily selected but were determined through manual iterative tuning. Various combinations were tested by monitoring the validation loss across folds during time-series cross-validation, and the configuration demonstrating optimal generalisation was selected. While automated tuning strategies such as grid search were not utilised due to computational constraints, the adopted approach enabled effective performance optimisation and ensured a transparent and reproducible modelling process. Model evaluation was conducted using the coefficient of determination (R2) across five regions, demonstrating consistent predictive accuracy.
Table 6, which summarises the performance metrics of the BiLSTM model for solar irradiance forecasting, highlights its high accuracy and consistent performance across regions, with notable variations in specific metrics. The model demonstrated exceptional performance in Yüreğir, achieving the highest accuracy, with a strong correlation (R = 0.9), the lowest MSE (0.01), and an RMSE of 0.12. A MAPE of 7.2% further confirms the model’s ability to capture solar irradiance patterns with minimal error in this region. Similarly, Fethiye achieved the highest correlation (R = 0.95) but with a slightly higher MSE (0.05) and RMSE (0.22) while maintaining a very low MAPE of 5.4%. These results indicate that the model is highly effective in forecasting solar irradiance in these regions, reflecting consistent and predictable patterns. In Adana, the model performed well, with a strong correlation (R = 0.84), but higher error metrics (MSE = 0.07, RMSE = 0.26) and a MAPE of 11.3% suggest some prediction variability. Isparta and Ulukışla displayed moderate correlations (R = 0.80), with comparable error metrics. Ulukışla had an MSE of 0.03, RMSE of 0.17, and MAPE of 12.1%, while Isparta showed an MSE of 0.07, RMSE of 0.26, and MAPE of 13.8%. These metrics indicate that, while the model captured temporal patterns reasonably well, some challenges remain in fully modelling the variability in these regions.
Overall, the BiLSTM model demonstrated robust predictive capabilities, excelling in regions with stable solar irradiance patterns, such as Yüreğir and Fethiye. The moderate performance in Adana, Isparta, and Ulukışla highlights the need for further refinements, such as optimising hyperparameters or incorporating additional input features, to improve forecasting accuracy in regions with more significant data variability.
The following figures (Figure 13, Figure 14, Figure 15, Figure 16 and Figure 17) depict the actual, test, and forecasted solar irradiance values obtained using the BiLSTM model from 2020 to 2026. The blue section illustrates the historical solar irradiance data from 2020 to 2023, providing a baseline for comparison. The red section represents the test phase, during which the BiLSTM model’s predictions were compared against actual data to assess its accuracy. Finally, the green section shows the forecasted solar irradiance values from 2023 to 2026, based on the model’s ability to capture bidirectional temporal patterns. These figures comprehensively evaluate the BiLSTM model’s capacity to predict future solar irradiance trends. The comparison between actual, test, and forecasted values highlights the model’s effectiveness in accurately forecasting solar irradiance levels across different regions.
The comparison between the actual and predicted values offers critical insights into the predictive accuracy of the BiLSTM model. According to the BiLSTM analysis, the forecasted annual average solar irradiance levels for 2026 are as follows: Yüreğir is expected to reach 4.87 kWh/m2/day, with an error rate of 7.2%, demonstrating a strong agreement between predictions and actual values. Ulukışla is forecasted at 4.70 kWh/m2/day, with a slightly higher error rate of 12.1%, indicating moderate variability in prediction accuracy. Fethiye is predicted to achieve 5.00 kWh/m2/day, with the lowest error rate of 5.4%, showcasing the model’s robustness in this region. Isparta is projected to record 4.80 kWh/m2/day, with an error rate of 13.8%, reflecting a more challenging prediction scenario than other regions. Finally, Adana is forecasted to have an annual average of 4.60 kWh/m2/day, with an error rate of 11.3%, signifying moderate accuracy. These results emphasise the BiLSTM model’s effectiveness in estimating annual solar irradiance averages while accounting for regional differences and associated prediction uncertainties.
Based on the BiLSTM analysis of solar irradiance patterns across the five regions—Yüreğir, Ulukışla, Fethiye, Adana, and Isparta—the model highlights consistent seasonal fluctuations, with each region demonstrating predictable peaks and troughs in solar irradiance. However, the forecast up to 2026 suggests a gradual decrease in solar irradiance levels for most areas, including Yüreğir, Ulukışla, Adana, and Isparta, while Fethiye is expected to maintain stable solar irradiance trends. This gradual decline in peak solar irradiance values may impact solar energy potential slightly, especially in regions like Yüreğir and Isparta, where the decrease is more noticeable. Despite this, the overall seasonal patterns remain intact, indicating that the areas will continue to experience regular, cyclical changes in solar irradiance. These results underline the importance of region-specific forecasting in understanding localised solar energy potential. Fethiye shows more robust and stable performance than the gradual decline observed in other areas.

4. Discussion

Solar radiation forecasting plays a crucial role in optimising renewable energy systems, enabling efficient energy management, and supporting sustainability goals. In regions like the Mediterranean, characterised by high solar energy potential and regional climatic variability, accurate forecasting methods are essential. This study utilised ANN, LSTM, and BiLSTM models to evaluate the spatio-temporal variations in solar radiation, highlighting their capabilities and limitations across diverse regions. A comparative analysis of the model performances across regions is presented in Figure 18.
ANN models were employed due to their ability to approximate nonlinear and complex interactions in solar irradiance data. However, in this study, their performance was moderate, with R values ranging from 0.45 in Ulukışla to 0.73 in Fethiye at a significance level of α = 0.05. This aligns with findings from prior research, which emphasised the adaptability of ANN models but noted their limitations in handling highly variable data [43,54,55]. The ANN model demonstrated its best performance in Fethiye (R = 0.73; MAPE = 8.2%), which can be attributed to the region’s relatively stable solar radiation patterns. Fethiye experiences significant seasonal variation in daily incident shortwave solar energy, with higher stability during summer. The brighter period of the year lasts for approximately 3.4 months, from May 10 to August 23, with an average daily incident shortwave energy per square meter above 7.2 kWh [56]. This seasonal stability likely contributes to the improved performance of the ANN model in this region. Conversely, the model exhibited higher errors in Ulukışla and Adana, with MAPE values of 23.3% and 24%, respectively, despite moderate R values (0.45 for Ulukışla and 0.64 for Adana). In Adana, these elevated errors may be attributed to complex meteorological factors, such as frequent transient cloud cover, high levels of urban and industrial pollution, and localised atmospheric instability. These factors introduce significant noise and irregular fluctuations into the solar irradiance data, challenging the ANN model’s capacity to generalise under such dynamically varying conditions.
LSTM models, designed to capture temporal dependencies in sequential data, demonstrated improved performance over ANN, with R values ranging from 0.61 in Ulukışla to 0.83 in Fethiye. These results align with prior studies highlighting the capability of classical LSTM to model intricate temporal patterns. While CNN-1D networks have outperformed LSTM in certain short-term solar irradiance forecasts, particularly at 10 min horizons, the robustness of LSTM layers in handling sequential data influenced by endogenous and exogenous factors remains well established. Supporting literature has shown that LSTM models provide reliable predictions in solar power forecasting due to their adaptability to dynamic time-series data [57,58,59]. In this study, improvements were observed across all regions when compared to ANN performance. For example, in Yüreğir, the R value increased modestly from 0.65 with ANN to 0.68 with LSTM, illustrating LSTM’s ability to adapt to sequential dependencies. Fethiye showed the highest R value of 0.83, accompanied by a low MAPE of 7.9%, reinforcing the model’s effectiveness in regions with more stable solar radiation trends. However, in Ulukışla, the LSTM model’s performance was relatively lower (R = 0.61), with a higher MAPE of 21.03%. This reduced accuracy may be attributed to the region’s complex topography and meteorological variability. Ulukışla is characterised by mountainous terrain and rapidly changing weather conditions, including frequent cloud cover and orographic effects, which introduce significant fluctuations in solar irradiance. Such dynamic and heterogeneous atmospheric patterns challenge the model’s ability to generalise across irregular time-series patterns, limiting its performance in this context.
BiLSTM models demonstrated the highest predictive accuracy across all regions in this study, achieving R values of 0.95 in Fethiye and 0.90 in Yüreğir. The bidirectional architecture of BiLSTM allows for it to utilise both past and future data points, significantly enhancing its capacity to model complex temporal dependencies. This advantage was especially evident in Fethiye, where stable solar radiation patterns facilitated the model’s superior performance. In Ulukışla, despite greater variability in solar irradiance, BiLSTM improved the R value to 0.80, markedly outperforming the ANN and LSTM models. However, the limitations of BiLSTM in highly variable regions such as Ulukışla and Adana warrant further investigation. The complex meteorological dynamics—characterised by rapid weather changes, orographic influences, and localised atmospheric disturbances—pose significant challenges that may not be fully addressed by a univariate BiLSTM approach. To enhance predictive performance in such areas, it is advisable to integrate additional meteorological variables, including temperature, wind speed, and humidity. Incorporating terrain features and increasing temporal resolution through hourly data could further improve sensitivity to local climatic variations. While BiLSTM achieved superior accuracy, it also introduces higher computational cost and longer training times due to its bidirectional structure, which may increase the risk of overfitting—especially in small or noisy datasets. Therefore, model complexity should be carefully balanced against available computational resources. Future research should focus on developing multivariate and hybrid models that combine BiLSTM with complementary techniques to effectively capture these complexities. The superiority of BiLSTM over other forecasting methods has been supported by studies focusing on regions with diverse climatic conditions, demonstrating its ability to capture intricate temporal dependencies and significantly improve prediction accuracy. For instance, Bektaş [38] highlighted the high accuracy of BiLSTM models in day-ahead hourly global horizontal irradiance forecasting, particularly in regions characterised by variable solar patterns. Similarly, region-specific studies in Mediterranean-like climates have demonstrated the effectiveness of BiLSTM in capturing complex temporal dependencies, leading to significant improvements in prediction accuracy. Alizadegan et al. [60] reported that BiLSTM demonstrated superior performance over LSTM in solar energy forecasting by more effectively capturing temporal dependencies, leading to improved accuracy in predicting energy output under varying climatic conditions. These findings align with the results of this study, further validating BiLSTM’s robustness in handling sequential data and its adaptability to varying solar radiation dynamics. Incorporating insights from regionally focused research underscores BiLSTM’s reliability for solar irradiance prediction tasks, particularly in geographically diverse regions like the Mediterranean. These findings emphasise the effectiveness of BiLSTM in regions with varying solar radiation dynamics, positioning it as a reliable tool for renewable energy forecasting tasks. Its performance underscores the importance of model selection tailored to the temporal and spatial characteristics of the data. An analysis of the predicted and observed solar radiation values reveals a damped oscillation phenomenon, particularly evident in ANN-based forecasts, where the amplitude of fluctuations decreases over time. This attenuation likely contributed to the performance differences among models, with ANN exhibiting lower accuracy in regions with more variable irradiance patterns, such as Ulukışla and Yüreğir, due to its limited ability to preserve high-frequency components in time-series predictions. A similar effect has been observed in climate model simulations investigating future solar irradiance variability. Sedlacek et al. [61], in their investigation of solar activity under CMIP6 scenarios, reported that reductions in solar forcing led to damped oscillations in long-term climate projections. Model constraints and preprocessing techniques may contribute to the gradual attenuation of oscillatory patterns. This provides insight into the comparatively weaker performance of ANN models in certain regions of this study. In contrast, the LSTM and BiLSTM models retained more of the underlying temporal structure, resulting in improved accuracy. The bidirectional structure of BiLSTM, which incorporates both past and future dependencies, further mitigated this effect, allowing for more effective modelling of sequential patterns.
To optimise solar irradiance forecasting in the Mediterranean region, prioritising BiLSTM models is recommended due to their superior adaptability and accuracy in capturing complex temporal dependencies. Prior research has demonstrated that including additional meteorological inputs substantially enhances predictive performance. For instance, Alharbi and Csala [62] employed a BiLSTM model augmented with wind speed and ambient temperature data, achieving a marked reduction in RMSE from 0.124 in the univariate setting to 0.078 in the multivariate framework, indicating a 37% improvement. Similarly, Alizamir et al. [63] integrated climatic variables into a wavelet-transformed LSTM architecture, reporting an RMSE decline from 3.42 MJ/m2 to 2.31 MJ/m2. Assaf et al. [64] further underscored the benefit of auxiliary inputs, noting RMSE reductions of approximately 10–15% when humidity and temperature were included. While such enhancements could not be implemented in the present study due to the unavailability of site-specific meteorological records, observed forecasting limitations in climatically volatile regions such as Ulukışla and Adana suggest that similar improvements might be realised through multivariate modelling. Future research should, therefore, pursue formal ablation studies to quantify the marginal contribution of each meteorological parameter and evaluate the efficacy of hybrid models—particularly those incorporating wavelet-based decomposition—to mitigate structural limitations, such as the damped oscillation effect observed in ANN-based forecasts. Incorporating higher temporal resolution (e.g., hourly intervals) and terrain-sensitive features may also enhance forecasting precision in topographically complex regions. These strategies are expected to support more accurate, resilient, and regionally transferable solar irradiance forecasting models across the Mediterranean basin.

5. Conclusions and Future Work

This study analysed spatio-temporal variations in solar irradiance across the Mediterranean region, comparing the performance of ANN, LSTM, and BiLSTM models in forecasting solar irradiance up to 2026. BiLSTM demonstrated the highest predictive accuracy, particularly in regions like Fethiye and Yüreğir, due to its ability to model temporal dependencies effectively. In contrast, ANN and LSTM showed moderate performance, with challenges in areas like Ulukışla and Adana, where data variability impacted accuracy. The findings highlight the importance of tailoring models to regional characteristics for optimised solar irradiance forecasting. Quantitatively, BiLSTM reduced the average RMSE by approximately 75% (from 0.83 to 0.21) and MAPE by 45% (from 18.2% to 10%) while increasing the average R value from 0.63 to 0.86 compared to the ANN model.
Future research may consider the integration of additional climatic and geographical features to improve predictive accuracy, particularly in regions exhibiting high spatio-temporal variability. Enhancing model robustness through the exploration of hybrid architectures or systematic hyperparameter optimisation is also recommended. Furthermore, extending the scope of analysis to encompass other renewable energy indicators could facilitate a more holistic approach to solar energy planning in the Mediterranean context. In addition, the incorporation of climate change scenarios—such as the Representative Concentration Pathway (RCP) 8.5—into solar radiation forecasting models would enable a more comprehensive understanding of long-term variability in solar resource availability. Recent studies [65,66] have demonstrated that scenario-based projections not only reveal region-specific trends in irradiance but also highlight the potential efficiency trade-offs arising from temperature increases. Accordingly, the fusion of deep learning-based forecasting frameworks with climate model simulations is expected to improve the resilience and adaptability of solar energy systems in response to future atmospheric and environmental shifts. Moreover, as daily-resolution data may obscure intra-day variability—an important factor for grid stability—future studies are encouraged to employ higher temporal resolution datasets (e.g., hourly measurements) to improve the responsiveness and operational relevance of solar irradiance forecasts.

Author Contributions

All authors contributed to the conception and design of the study. B.İ. and U.Ş. performed material preparation, data collection, and analysis. B.İ. wrote the manuscript’s first draft, which was later critically reviewed and edited by A.T., Z.A. and R.H. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author(s).

Acknowledgments

This study is supported by the North Atlantic Treaty Organization (NATO) Science for Peace and Security (SPS) Multi-Year Project number G5970, named Cube4EnvSec: ‘Big Earth Data-cube Analytics for Transnational Security and Environment,’ https://cube4envsec.org/ (accessed on 1 June 2022).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Kumari, P.; Toshniwal, D. Deep learning models for solar irradiance forecasting: A comprehensive review. J. Clean. Prod. 2021, 318, 128566. [Google Scholar] [CrossRef]
  2. Shadab, A.; Ahmad, S.; Said, S. Comparative evaluation of ARIMA and ARCGIS solar analyst for spatial insolation forecasts-A case study. Water Energy Int. 2024, 67, 29–36. [Google Scholar]
  3. Ministry of Energy and Natural Resources (MENR). Energy Statistics of Turkey in 2023. Available online: http://www.enerji.gov.tr (accessed on 6 February 2025).
  4. Incecik, S.; Sakarya, S.; Tilev, S.; Kahraman, A.; Aksoy, B.; Calıska, E.; Topçu, S.; Kâhya, C.; Odman, M.T. Evaluation of WRF parameterizations for global horizontal irradiation forecasts: A study for Turkey. Atmósfera 2019, 32, 143–158. [Google Scholar] [CrossRef]
  5. Belmahdi, B.; Louzazni, M.; El Bouardi, A. Comparative optimization of global solar irradiance forecasting using machine learning and time series models. Environ. Sci. Pollut. Res. 2022, 29, 14871–14888. [Google Scholar] [CrossRef] [PubMed]
  6. Son, W.; Lee, Y.R. Day-Ahead Prediction of PV Power Output: A One-Year Case Study at Changwon in South Korea. J. Electr. Eng. Technol. 2024, 20, 71–79. [Google Scholar] [CrossRef]
  7. Singh, P.K.; Saraswat, A.; Gupta, Y.; Goyal, S.K.; Gupta, Y. A Comparative Study of Deep Learning Methods for Short-Term Solar Radiation Forecasting. In Flexible Electronics for Electric Vehicles; Springer Nature Singapore: Singapore, 2022; pp. 565–575. [Google Scholar]
  8. Yu, H.; Jiang, S.; Chen, M.; Wang, M.; Shi, R.; Li, S.; Zhan, C. Machine learning models for daily net radiation prediction across different climatic zones of China. Sci. Rep. 2024, 14, 20454. [Google Scholar] [CrossRef] [PubMed]
  9. Guven, D. Analysing the Determinants of Surface Solar Radiation with Tree-Based Machine Learning Methods: Case of Istanbul. Pure Appl. Geophys. 2024, 181, 1633–1659. [Google Scholar] [CrossRef]
  10. Pérez-Rodríguez, S.A.; Álvarez-Alvarado, J.M.; Romero-González, J.A.; Aviles, M.; Eileen, M.R.A.; Carlos, F.S.; Rodríguez-Reséndiz, J. Metaheuristic algorithms for solar radiation prediction: A systematic analysis. IEEE Access 2024, 12, 100134–100151. [Google Scholar] [CrossRef]
  11. Yahiaoui, S.; Assas, O. Comparison of solar irradiance models using meteorological parameters. Energy Syst. 2024, 15, 863–897. [Google Scholar] [CrossRef]
  12. Minuzzi, F.C.; Farina, L. A deep learning approach to predict significant wave height using long short-term memory. Ocean Model. 2023, 181, 102151. [Google Scholar] [CrossRef]
  13. Haider, S.A.; Sajid, M.; Sajid, H.; Uddin, E.; Ayaz, Y. Deep learning and statistical methods for short-and long-term solar irradiance forecasting for Islamabad. Renew. Energy 2022, 198, 51–60. [Google Scholar] [CrossRef]
  14. Brahma, B.; Wadhvani, R. Solar irradiance forecasting based on deep learning methodologies and multi-site data. Symmetry 2020, 12, 1830. [Google Scholar] [CrossRef]
  15. Ryu, S.; Noh, J.; Kim, H. Deep neural network based demand side short term load forecasting. Energies 2016, 10, 3. [Google Scholar] [CrossRef]
  16. İşler, B.; Şener, U.; Tokgözlü, A.; Aslan, Z.; Baumann, P. Prediction of Solar Energy Potential with Machine Learning and Deep Learning Models. In Mathematical Modeling, Applied Analysis and Computational Methods, Proceedings of the ICAIM 2023, Greater Noida, India, 24–26 March 2023; Springer Proceedings in Mathematics & Statistics; Alam, K., Khan, A., Singh, R.C., Karaca, Y., Eds.; Springer: Singapore, 2025; Volume 482. [Google Scholar] [CrossRef]
  17. Chodakowska, E.; Nazarko, J.; Nazarko, Ł.; Rabayah, H.S. Solar Radiation Forecasting: A Systematic Meta-Review of Current Methods and Emerging Trends. Energies 2024, 17, 3156. [Google Scholar] [CrossRef]
  18. Ghenai, C.; Ahmad, F.F.; Rejeb, O. Artificial neural network-based models for short term forecasting of solar PV power output and battery state of charge of solar electric vehicle charging station. Case Stud. Therm. Eng. 2024, 61, 105152. [Google Scholar] [CrossRef]
  19. Terregrossa, S.J.; Şener, U. Employing a generalized reduced gradient algorithm method to form combinations of steel price forecasts generated separately by ARIMA-TF and ANN models. Cogent Econ. Financ. 2023, 11, 2169997. [Google Scholar] [CrossRef]
  20. Kong, W.; Dong, Z.Y.; Jia, Y.; Hill, D.J.; Xu, Y.; Zhang, Y. Short-term residential load forecasting based on LSTM recurrent neural network. IEEE Trans. Smart Grid 2017, 10, 841–851. [Google Scholar] [CrossRef]
  21. Elsaraiti, M.; Merabet, A. A comparative analysis of the arima and lstm predictive models and their effectiveness for predicting wind speed. Energies 2021, 14, 6782. [Google Scholar] [CrossRef]
  22. Hou, J.; Wang, Y.; Zhou, J.; Tian, Q. Prediction of hourly air temperature based on CNN–LSTM. Geomat. Nat. Hazards Risk 2022, 13, 1962–1986. [Google Scholar] [CrossRef]
  23. Shi, X.; Chen, Z.; Wang, H.; Yeung, D.Y.; Wong, W.K.; Woo, W.C. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. Adv. Neural Inf. Process. Syst. 2015, 28, 802–810. [Google Scholar]
  24. Gamboa-Villafruela, C.J.; Fernández-Alvarez, J.C.; Márquez-Mijares, M.; Pérez-Alarcón, A.; Batista-Leyva, A.J. Convolutional lstm architecture for precipitation nowcasting using satellite data. Environ. Sci. Proc. 2021, 8, 33. [Google Scholar]
  25. Yang, A. Big data-driven corporate financial forecasting and decision support: A study of CNN-LSTM machine learning models. Front. Appl. Math. Stat. 2025, 11, 1566078. [Google Scholar] [CrossRef]
  26. Mohanasundaram, V.; Rangaswamy, B. Photovoltaic solar energy prediction using the seasonal-trend decomposition layer and ASOA optimized LSTM neural network model. Sci. Rep. 2025, 15, 4032. [Google Scholar] [CrossRef] [PubMed]
  27. Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735–1780. [Google Scholar] [CrossRef] [PubMed]
  28. Greff, K.; Srivastava, R.K.; Koutník, J.; Steunebrink, B.R.; Schmidhuber, J. LSTM: A Search Space Odyssey. IEEE Trans. Neural Netw. Learn. Syst. 2017, 28, 2222–2232. [Google Scholar] [CrossRef] [PubMed]
  29. Gomathi, S.; Kannan, E.; Belinda, M.C.M.; Giri, J.; Nagaraju, V.; Kumar, J.A.; Praveenkumar, T.R. Solar energy prediction with synergistic adversarial energy forecasting system (Solar-SAFS): Harnessing advanced hybrid techniques. Case Stud. Therm. Eng. 2024, 63, 105197. [Google Scholar] [CrossRef]
  30. Yildirim, A.; Bilgili, M.; Kara, O. Deep learning approach for one-hour ahead forecasting of solar radiation in different climate regions. Int. J. Green Energy 2024, 21, 2984–3000. [Google Scholar] [CrossRef]
  31. Campos, F.D.; Sousa, T.C.; Barbosa, R.S. Short-Term Forecast of Photovoltaic Solar Energy Production Using LSTM. Energies 2024, 17, 2582. [Google Scholar] [CrossRef]
  32. Ehteram, M.; Nia, M.A.; Panahi, F.; Farrokhi, A. Read-First LSTM model: A new variant of long short-term memory neural network for predicting solar radiation data. Energy Convers. Manag. 2024, 305, 118267. [Google Scholar] [CrossRef]
  33. Olcay, K.; Tunca, S.G.; Özgür, M.A. Forecasting and performance analysis of energy production in solar power plants using long short-term memory (LSTM) and random forest models. IEEE Access 2024, 12, 103299–103312. [Google Scholar] [CrossRef]
  34. Fakhriza, I. Perbandingan Prediksi Radiasi Sinar Matahari Menggunakan Algoritma Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM) dan Deep Q-Network (DQN). Ph.D. Dissertation, Universitas Sumatera Utara, Medan, Indonesia, 2024. [Google Scholar]
  35. Peng, T.; Zhang, C.; Zhou, J.; Nazir, M.S. An integrated framework of Bi-directional long-short term memory (BiLSTM) based on sine cosine algorithm for hourly solar radiation forecasting. Energy 2021, 221, 119887. [Google Scholar] [CrossRef]
  36. Bou-Rabee, M.A.; Naz, M.Y.; Albalaa, I.E.; Sulaiman, S.A. BiLSTM Network-Based Approach for Solar Irradiance Forecasting in Continental Climate Zones. Energies 2022, 15, 2226. [Google Scholar] [CrossRef]
  37. Li, C.; Zhang, Y.; Zhao, G.; Ren, Y. Hourly solar irradiance prediction using deep BiLSTM network. Earth Sci. Inform. 2021, 14, 299–309. [Google Scholar] [CrossRef]
  38. Bektaş, S.Ç.; Altaş, I.H. DWT-BILSTM-based models for day-ahead hourly global horizontal solar irradiance forecasting. Neural Comput. Appl. 2024, 36, 13243–13253. [Google Scholar] [CrossRef]
  39. Sanganaboina, A.B.; Ruttala, S.; Mandadapu, H.; Kanigiri, S.V.U.M.; Kumar, S.D.; Kumar, S.V.P. Prediction of Solar Panel Maintenance using BiLSTM. In Proceedings of the 2024 3rd International Conference on Applied Artificial Intelligence and Computing (ICAAIC), Salem, India, 5–7 June 2024; IEEE: Piscataway, NJ, USA, 2024; pp. 1880–1886. [Google Scholar]
  40. TSMS—Turkish State Meteorological Service. Available online: https://www.mgm.gov.tr/ (accessed on 1 June 2024).
  41. El-Amarty, N.; Marzouq, M.; El Fadili, H.; Bennani, S.D.; Ruano, A. A comprehensive review of solar irradiation estimation and forecasting using artificial neural networks: Data, models and trends. Environ. Sci. Pollut. Res. 2023, 30, 5407–5439. [Google Scholar] [CrossRef] [PubMed]
  42. Yuzer, E.O.; Bozkurt, A. Instant solar irradiation forecasting for solar power plants using different ANN algorithms and network models. Electr. Eng. 2024, 106, 3671–3689. [Google Scholar] [CrossRef]
  43. Rahman, S.; Rahman, S.; Haque, A.B. Prediction of solar radiation using artificial neural network. J. Phys. Conf. Ser. 2021, 1767, 012041. [Google Scholar] [CrossRef]
  44. Jailani, N.L.M.; Dhanasegaran, J.K.; Alkawsi, G.; Alkahtani, A.A.; Phing, C.C.; Baashar, Y.; Tiong, S.K. Investigating the power of LSTM-based models in solar energy forecasting. Processes 2023, 11, 1382. [Google Scholar] [CrossRef]
  45. Yildirim, A.; Bilgili, M.; Ozbek, A. One-hour-ahead solar radiation forecasting by MLP, LSTM, and ANFIS approaches. Meteorol. Atmos. Phys. 2023, 135, 10. [Google Scholar] [CrossRef]
  46. Hoyos-Gómez, L.S.; Ruiz-Muñoz, J.F.; Ruiz-Mendoza, B.J. Short-term forecasting of global solar irradiance with incomplete data. arXiv 2021, arXiv:2106.06868. [Google Scholar]
  47. Demir, V.; Citakoglu, H. Forecasting of solar radiation using different machine learning approaches. Neural Comput. Appl. 2023, 35, 887–906. [Google Scholar] [CrossRef]
  48. Abotaleb, M.; Dutta, P.K. Optimizing bidirectional long short-term memory networks for univariate time series forecasting: A comprehensive guide. In Hybrid Information Systems: Non-Linear Optimization Strategies with Artificial Intelligence; IEEE: Piscataway, NJ, USA, 2024; p. 443. [Google Scholar]
  49. He, Z.; Zhang, X.; Li, M.; Wang, S.; Xiao, G. A novel solar radiation forecasting model based on time series imaging and bidirectional long short-term memory network. Energy Sci. Eng. 2024, 12, 4876–4893. [Google Scholar] [CrossRef]
  50. Zameer, A.; Jaffar, F.; Shahid, F.; Muneeb, M.; Khan, R.; Nasir, R. Short-term solar energy forecasting: Integrated computational intelligence of LSTMs and GRU. PLoS ONE 2023, 18, e0285410. [Google Scholar] [CrossRef] [PubMed]
  51. Goswami, S.; Malakar, S.; Ganguli, B.; Chakrabarti, A. A novel transfer learning-based short-term solar forecasting approach for India. Neural Comput. Appl. 2022, 34, 16829–16843. [Google Scholar] [CrossRef]
  52. Şener, U.; Terregrossa, S.J. A Transcendental LASSO Function for Combining Machine Learning and Statistical Model Forecasts. SAGE Open 2024, 14, 21582440241262695. [Google Scholar] [CrossRef]
  53. Şener, U.; Kılıç, B.I.; Tokgözlü, A.; Aslan, Z. Prediction of Wind Speed by Using Machine Learning. In Computational Science and Its Applications—ICCSA 2023 Workshops, Proceedings of the ICCSA 2023, Athens, Greece, 3–6 July 2023; Gervasi, O., Ed.; Lecture Notes in Computer Science; Springer: Cham, Switzerland, 2023; Volume 14104. [Google Scholar] [CrossRef]
  54. Malik, P.; Gehlot, A.; Singh, R.; Gupta, L.R.; Thakur, A.K. A review on ANN based model for solar radiation and wind speed prediction with real-time data. Arch. Comput. Methods Eng. 2022, 29, 3183–3201. [Google Scholar] [CrossRef]
  55. Ledmaoui, Y.; El Maghraoui, A.; El Aroussi, M.; Saadane, R.; Chebak, A.; Chehri, A. Forecasting solar energy production: A comparative study of machine learning algorithms. Energy Rep. 2023, 10, 1004–1012. [Google Scholar] [CrossRef]
  56. WeatherSpark. 2024. Available online: https://weatherspark.com/y/95926/Average-Weather-in-Fethiye-Turkey-Year-Round#google_vignette (accessed on 1 February 2025).
  57. Marinho, F.P.; Rocha, P.A.; Neto, A.R.; Bezerra, F.D. Short-term solar irradiance forecasting using CNN-1D, LSTM, and CNN-LSTM deep neural networks: A case study with the Folsom (USA) dataset. J. Sol. Energy Eng. 2023, 145, 041002. [Google Scholar] [CrossRef]
  58. Khan, S.Z.; Muzammil, N.; Ghafoor, S.; Khan, H.; Zaidi, S.M.H.; Aljohani, A.J.; Aziz, I. Quantum long short-term memory (QLSTM) vs. classical LSTM in time series forecasting: A comparative study in solar power forecasting. Front. Phys. 2024, 12, 1439180. [Google Scholar] [CrossRef]
  59. AlKandari, M.; Ahmad, I. Solar power generation forecasting using ensemble approach based on deep learning and statistical methods. Appl. Comput. Inform. 2024, 20, 231–250. [Google Scholar] [CrossRef]
  60. Alizadegan, H.; Radmehr, A.; Karimi, H.; Ilani, M.A. Solar Energy Production Forecasting: A Comparative Study of Bi-LSTM, LSTM, XGBoost Models with Activation Function Analysis. Preprint 2024. [Google Scholar] [CrossRef]
  61. Sedlacek, J.; Sukhodolov, T.; Egorova, T.; Karagodin-Doyennel, A.; Rozanov, E. Future climate under CMIP6 solar activity scenarios. Earth Space Sci. 2023, 10, e2022EA002783. [Google Scholar] [CrossRef]
  62. Alharbi, F.R.; Csala, D. Wind speed and solar irradiance prediction using a bidirectional long short-term memory model based on neural networks. Energies 2021, 14, 6501. [Google Scholar] [CrossRef]
  63. Alizamir, M.; Shiri, J.; Fard, A.F.; Kim, S.; Gorgij, A.D.; Heddam, S.; Singh, V.P. Improving the accuracy of daily solar radiation prediction by climatic data using an efficient hybrid deep learning model: Long short-term memory (LSTM) network coupled with wavelet transform. Eng. Appl. Artif. Intell. 2023, 123, 106199. [Google Scholar] [CrossRef]
  64. Assaf, A.M.; Haron, H.; Abdull Hamed, H.N.; Ghaleb, F.A.; Qasem, S.N.; Albarrak, A.M. A review on neural network based models for short term solar irradiance forecasting. Appl. Sci. 2023, 13, 8332. [Google Scholar] [CrossRef]
  65. MedECC. First Mediterranean Assessment Report: Climate and Environmental Change in the Mediterranean Basin—Current Situation and Risks for the Future; MedECC: Marseille, France, 2020. [Google Scholar]
  66. Hou, X.; Wild, M.; Folini, D.; Kazadzis, S.; Wohland, J. Climate change impacts on solar power generation and its spatial variability in Europe based on CMIP6. Earth Syst. Dyn. 2021, 12, 1099–1113. [Google Scholar] [CrossRef]
Figure 1. Study area.
Figure 1. Study area.
Sustainability 17 06696 g001
Figure 2. Flow chart of the machine learning models used to estimate solar irradiance.
Figure 2. Flow chart of the machine learning models used to estimate solar irradiance.
Sustainability 17 06696 g002
Figure 3. ANN model outputs for solar irradiance in Yüreğir. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 3. ANN model outputs for solar irradiance in Yüreğir. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g003
Figure 4. ANN model outputs for solar irradiance in Ulukışla. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 4. ANN model outputs for solar irradiance in Ulukışla. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g004
Figure 5. ANN model outputs for solar irradiance in Fethiye. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 5. ANN model outputs for solar irradiance in Fethiye. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g005
Figure 6. ANN model outputs for solar irradiance in Adana. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 6. ANN model outputs for solar irradiance in Adana. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g006
Figure 7. ANN model outputs for solar irradiance in Isparta. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 7. ANN model outputs for solar irradiance in Isparta. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g007
Figure 8. LSTM model outputs for solar irradiance in Yüreğir. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 8. LSTM model outputs for solar irradiance in Yüreğir. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g008
Figure 9. LSTM model outputs for solar irradiance in Ulukışla. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 9. LSTM model outputs for solar irradiance in Ulukışla. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g009
Figure 10. LSTM model outputs for solar irradiance in Fethiye. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 10. LSTM model outputs for solar irradiance in Fethiye. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g010
Figure 11. LSTM model outputs for solar irradiance in Adana. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 11. LSTM model outputs for solar irradiance in Adana. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g011
Figure 12. LSTM model outputs for solar irradiance in Isparta. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 12. LSTM model outputs for solar irradiance in Isparta. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g012
Figure 13. BiLSTM model outputs for solar irradiance in Yüreğir. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 13. BiLSTM model outputs for solar irradiance in Yüreğir. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g013
Figure 14. BiLSTM model outputs for solar irradiance in Ulukışla. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 14. BiLSTM model outputs for solar irradiance in Ulukışla. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g014
Figure 15. BiLSTM model outputs for solar irradiance in Fethiye. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 15. BiLSTM model outputs for solar irradiance in Fethiye. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g015
Figure 16. BiLSTM model outputs for solar irradiance in Adana. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 16. BiLSTM model outputs for solar irradiance in Adana. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g016
Figure 17. BiLSTM model outputs for solar irradiance in Isparta. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Figure 17. BiLSTM model outputs for solar irradiance in Isparta. The blue section indicates the training period (2020–2023), red denotes the testing period, and green represents the forecasted values (2023–2026).
Sustainability 17 06696 g017
Figure 18. Heatmap comparison of ANN-NARX, LSTM, and BiLSTM model performance metrics for solar irradiance forecasting across regions.
Figure 18. Heatmap comparison of ANN-NARX, LSTM, and BiLSTM model performance metrics for solar irradiance forecasting across regions.
Sustainability 17 06696 g018
Table 1. Summary of related work.
Table 1. Summary of related work.
RefKey FindingsResults and Data Used
[1]Demonstrates the need for advanced predictive models capable of capturing complex temporal and spatial dependencies in solar irradiance forecasting.The proposed models were validated under diverse temporal and spatial conditions using comprehensive datasets.
[2]The challenges posed by geographical heterogeneity and seasonal cycles in solar energy forecasting were addressed.Simulated results under diverse weather scenarios to account for geographical and seasonal variability.
[5]Traditional ARIMA methods were applied but were limited in capturing the nonlinear characteristics of solar irradiance data.RMSE and MAE metrics were used to evaluate historical data from rooftop solar PV systems.
[6]Statistical methods demonstrated interpretability but were limited in modelling dynamic solar patterns.Traditional statistical benchmarks were utilised, with metrics including R-squared and MAPE.
[10]ANNs improved predictive accuracy by capturing nonlinear relationships in solar irradiance data.RMSE and MAE metrics were applied to multi-variable datasets from solar irradiance systems.
[11]Limitations of ANNs in capturing long-term dependencies were identified, suggesting advanced architectures for sequential tasks.The integration of ANNs with hybrid models was proposed to enhance long-term forecasting accuracy.
[30]LSTM models demonstrated higher predictive accuracy in learning nonlinear and sequential patterns in solar irradiance data.LSTM models, trained on historical solar data, achieved superior accuracy, reflected in reduced RMSE values.
[31]BiLSTM models outperformed other methods in capturing intricate temporal patterns and bidirectional dependencies.BiLSTM models, tested on dynamic datasets, achieved lower MAE and RMSE compared to traditional methods.
[38]BiLSTM effectively modelled complex temporal patterns in solar irradiance forecasting with high accuracy and reliability.BiLSTM was proposed for day-ahead forecasting, using hourly solar irradiance data across different climatic zones.
Table 2. Monthly average solar irradiance levels (kWh/m2/day) across selected regions in the Mediterranean (2020–2023).
Table 2. Monthly average solar irradiance levels (kWh/m2/day) across selected regions in the Mediterranean (2020–2023).
MonthsYüreğirUlukışlaFethiyeAdanaIsparta
Jan2.402.142.262.252.79
Feb3.313.393.373.133.76
Mar4.373.614.514.114.52
Apr5.715.115.975.305.33
May6.926.306.896.586.00
Jun7.196.777.706.696.72
Jul7.517.627.756.847.75
Aug6.616.296.725.876.82
Sep5.575.645.735.155.61
Oct4.104.054.043.764.13
Nov2.752.692.682.602.78
Dec2.212.062.092.111.92
Table 3. Annual average solar irradiance levels (kWh/m2/day) across regions (2020–2023).
Table 3. Annual average solar irradiance levels (kWh/m2/day) across regions (2020–2023).
YearsYüreğirUlukışlaFethiyeAdanaIsparta
20204.794.534.894.535.14
20215.094.815.074.594.87
20224.864.645.054.554.82
20234.854.614.944.504.67
Table 4. Performance metrics of ANN NARX model for solar irradiance forecasting across areas.
Table 4. Performance metrics of ANN NARX model for solar irradiance forecasting across areas.
AreaRMSERMSEMAPE
Yüreğir0.650.590.7616.8
Ulukışla0.451.41.223.3
Fethiye0.730.30.618.2
Adana0.640.520.7424
Isparta0.690.670.8217.6
Table 5. Performance metrics of LSTM model for solar irradiance forecasting across areas.
Table 5. Performance metrics of LSTM model for solar irradiance forecasting across areas.
AreaRMSERMSEMAPE
Yüreğir0.680.520.7215.1
Ulukışla0.611.081.1721.03
Fethiye0.830.560.397.9
Adana0.740.820.6822.3
Isparta0.740.780.6215.4
Table 6. Performance metrics of BiLSTM model for solar irradiance forecasting across areas.
Table 6. Performance metrics of BiLSTM model for solar irradiance forecasting across areas.
AreaRMSERMSEMAPE
Yüreğir0.90.010.127.2
Ulukışla0.80.030.1712.1
Fethiye0.950.050.225.4
Adana0.840.070.2611.3
Isparta0.800.070.2613.8
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

İşler, B.; Şener, U.; Tokgözlü, A.; Aslan, Z.; Heise, R. Spatio-Temporal Variation in Solar Irradiance in the Mediterranean Region: A Deep Learning Approach. Sustainability 2025, 17, 6696. https://doi.org/10.3390/su17156696

AMA Style

İşler B, Şener U, Tokgözlü A, Aslan Z, Heise R. Spatio-Temporal Variation in Solar Irradiance in the Mediterranean Region: A Deep Learning Approach. Sustainability. 2025; 17(15):6696. https://doi.org/10.3390/su17156696

Chicago/Turabian Style

İşler, Buket, Uğur Şener, Ahmet Tokgözlü, Zafer Aslan, and Rene Heise. 2025. "Spatio-Temporal Variation in Solar Irradiance in the Mediterranean Region: A Deep Learning Approach" Sustainability 17, no. 15: 6696. https://doi.org/10.3390/su17156696

APA Style

İşler, B., Şener, U., Tokgözlü, A., Aslan, Z., & Heise, R. (2025). Spatio-Temporal Variation in Solar Irradiance in the Mediterranean Region: A Deep Learning Approach. Sustainability, 17(15), 6696. https://doi.org/10.3390/su17156696

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop