1. Introduction
The global shift toward renewable energy sources has made wind power a key component in achieving sustainable energy goals. By 2030, Latvia aims to expand wind capacity to 800 MW, aligning with broader European targets for clean energy. On a global scale, the combined capacity of onshore and offshore wind energy is expected to reach approximately 352 GW by 2030. This growth supports the European Green Deal, which aims to make Europe the first climate-neutral continent by 2050. Over 75% of the EU’s greenhouse gas emissions are linked to energy production and consumption. Transforming the energy system is therefore central to achieving climate neutrality. Wind energy stands out as a preferred solution due to its scalability, cost competitiveness, and potential to create jobs. However, as wind penetration in the energy mix increases, so does the need for accurate short-term forecasting. Reliable wind power forecasts are essential for grid integration, operational planning, market participation, and system stability. Despite advancements, short-term wind forecasting remains a complex challenge. However, its variable and unpredictable nature presents significant challenges for maintaining power grid stability, ensuring reliable energy dispatching, and ensuring the consistent operation of wind power plants. Fluctuations in wind speed can lead to substantial variations in turbine output, introducing vulnerabilities in energy systems. Accurate wind power forecasting reduces grid operation costs and enhances system reliability. Yet, forecasting remains complex due to factors such as climate variability, seasonal changes, and the intermittent nature of wind. Despite advancements in machine learning and predictive modelling, existing approaches often fail to deliver the necessary accuracy, particularly in regions with low or inconsistent wind speeds. Addressing these challenges requires improved wind speed prediction methods tailored to specific site conditions, ensuring the reliable and efficient integration of wind energy into modern power systems. Wind power forecasting depends on forecast horizons, geographical conditions, climatic variability, testing period, normalisation, and pre-processing techniques.
The main meteorological and environment factors affecting wind energy yield are terrain features surrounding site locations. Natural barriers (e.g., hills and mountains) and open spaces (e.g., coastal regions and valleys) significantly affect wind flows. Temperature and relative humidity influence air density and have some effect on wind energy yield, while forested areas induce higher turbulence on the air masses. Depending on the intended application, wind power forecasting is conducted over various time intervals (see
Table 1). In timeframes ranging from milliseconds to several minutes, forecasts can be used for active turbine control and real-time management of the entire system’s operation. This type of forecasting is commonly referred to as ultra-short-term forecasting. Short-term forecasts (from 1 h to 72 h) are essential for power dispatch planning, electricity supply scheduling, grid operation management, and electricity trading on the market. Moreover, such forecasts are critically crucial for making decisions about the use of conventional power plants and for optimising their operating modes. Medium-term forecasts, spanning a range of 3 to 7 days, are essential for the routine maintenance of wind power plants (WPPs), the overall energy system infrastructure, and for planning energy storage. For longer time intervals, forecasts are intended for the operational planning of wind farms or conventional power plants. Forecast uncertainties generally increase with the length of the forecasting horizon.
A wide range of regression models have been developed and applied in wind turbine energy forecasting to address the diverse and often complex relationships within wind datasets. Linear regression, as the most fundamental approach, predicts target variables using a straight line and is valued for its simplicity and interpretability. However, it cannot capture non-linear patterns in the data [
1]. More advanced linear models, such as Ridge and Lasso regression [
2,
3], incorporate regularisation techniques—L2 and L1, respectively—to prevent overfitting and perform automatic feature selection. At the same time, ElasticNet balances the strengths of both methods [
4]. Support vector regression (SVR) offers high accuracy for forecasting wind turbine energy yield by effectively modelling complex, non-linear relationships [
5]. Non-linear and ensemble methods, including decision tree, random forest, extra trees [
6,
7,
8], and various gradient boosting algorithms (such as XGBoost, CatBoost, and LightGBM) [
9,
10], further enhance predictive performance by capturing nonlinearity and reducing model bias and variance. These ensemble models are particularly effective in addressing uncertainties, handling multicollinearity, and improving computational efficiency, which are crucial for real-time and large-scale wind energy applications. In addition, specific algorithms, such as K-nearest neighbours (k-NN), are straightforward to implement and are particularly useful for geo-imputation of missing wind data, despite typically having lower accuracy compared to other models [
11]. Modern boosting methods, such as histogram-based gradient boosting and AdaBoost, offer further speed and robustness. At the same time, CatBoost’s capability to handle categorical features and integrate advanced pre-processing steps makes it highly suitable for complex wind datasets [
12,
13].
Various regression models have been explored to achieve accurate predictions, with advancements in machine learning (ML) and deep learning (DL) techniques. This analysis compares the performance of various regression models, with a focus on their predictive accuracy for wind turbine energy yield (
Table 2).
Unlike many previous studies that fuse multiple data sources into a single input set, this research conducts a comparative analysis of short-term wind energy forecasting accuracy using three distinct datasets: hub-height sensor measurements from wind turbines, meteorological observations from the Latvian Environment, Geology and Meteorology Centre (LEGMC), and the NORA 3 reanalysis dataset. By independently evaluating the performance of various machine learning models on each dataset, the study assesses the relative predictive power of each data source, offering valuable insights into the most effective inputs for short-term forecasting in the Baltic region. This study aims to analyse existing forecasting models and to identify more accurate time series prediction approaches suited to current conditions. Accurate forecasting requires the evaluation of atmospheric conditions, with particular attention to key variables such as wind speed and wind direction at hub height. However, obtaining reliable long-term observational data often necessitates the use of expensive equipment and extended measurement campaigns. In practice, wind turbine operators frequently have access only to short-term data, which presents additional challenges for precise wind forecasting and resource assessment. To address these limitations, this research compares data from multiple sources. The forecasting framework incorporates operational data from two wind turbines in Latvia, meteorological inputs from the NORA3 reanalysis dataset, sensor measurements from the turbines, and data provided by the Latvian Environment, Geology and Meteorology Centre (LEGMC). By integrating and evaluating these diverse data sources, the study aims to improve the reliability and accuracy of short-term wind power forecasting under real-world conditions.
2. Materials and Methods
The selected wind turbines are located near Ainaži, a coastal town in the northern part of Latvia, close to the Estonian border (see
Figure 1). Geographically, Ainaži lies along the eastern coast of the Gulf of Riga, offering favourable wind conditions due to the unobstructed airflow from the sea.
The region is characterised by flat coastal terrain with low vegetation and minimal topographical disruptions, contributing to steady wind profiles and reduced turbulence.
Meteorologically, Ainaži experiences a temperate maritime climate, with relatively mild winters and cool summers. The average annual wind speed in the region typically ranges between 6 and 8 m per second, making it suitable for wind energy production. The proximity to the sea also leads to more consistent wind patterns, especially during autumn and winter. However, seasonal variability and occasional coastal storms must be considered when modelling wind behaviour and forecasting energy output.
In the investigated location, two Vestas V39 wind turbines were installed, each rated for 500 kW of power. The V39 model, developed by Vestas Wind Systems, features a 42 m rotor diameter, pitch-regulated control, and an asynchronous generator.
Access to long-term and high-quality observational data is essential for machine learning models to effectively capture seasonal patterns, identify anomalies, and improve generalisation for future predictions. However, long-term wind observation records in Latvia—especially at hub height—are limited or completely absent, posing a challenge for training robust forecasting models. In this study, three independent datasets were integrated to enhance the accuracy of wind power prediction.
Meteorological variables in the study area were analysed using the NORA3 dataset [
24], a high-resolution (3 km) regional atmospheric reanalysis developed by the Norwegian Meteorological Institute. The dataset was downloaded in the form of several CSV files for years 2022 to 2024 and concatenated into a continuous dataset. It includes parameters such as temperature, humidity, pressure at 2 m, wind speed, gusts, and direction at both 10 and 50 m above sea level.
Figure 2a presents an analysis of wind speed distribution at a height of 50 m. The Weibull distribution closely approximates the observed wind speed histogram, with the highest frequency occurring at around 6–7 m/s. The wind rose (
Figure 2b) illustrates the distribution of wind direction and speed at an altitude of 50 m, based on data collected from 2022 to 2024.
Dominant wind directions are identified as southwest (SW), west–southwest (WSW), and south–southwest (SSW), collectively representing the highest percentage of wind occurrences. Most wind events fall within moderate speed categories (5–15 m/s), while higher wind speeds (15–20 m/s) occur less frequently and primarily from southwestern sectors. Less regularly observed winds are from northern and northeastern directions.
An extended period starting from the year 1999 to 2024 of the NORA3 dataset was used to evaluate impact of seasonality and trends in the environmental data. Average wind speeds have little change over these years, while dispersion has slightly increased. These changes currently manifest as stronger gusts compared to 25 years ago. However, this study focuses on short-term forecasting and on the limited datasets, thus these long-term trends are not considered in detail. The seasonality was found as expected—stronger winds during winter seasons—and was incorporated indirectly as a timestamp feature.
The second dataset was obtained from the Latvian Environment, Geology and Meteorology Centre [
25]. Offering site-specific historical weather observations that reflect local climate trends and seasonal variations, it includes parameters such as temperature, wind speed, gusts, and direction at an altitude of 6 m. The dataset was downloaded from the LEGMC web page in the form of monthly tables for each parameter (CSV files) and merged into a continuous dataset.
The third dataset originated from the wind turbine operator and included high-resolution wind data recorded by the Nacelle Sensor. This turbine-level data captured real-time wind conditions directly at the hub, making it especially valuable for operational forecasting. The dataset also contained detailed information about the actual energy production of the turbine. The dataset was provided in the form of several MS Excel (*.xlsx) files covering various periods and was parsed and concatenated into a continuous dataset.
As all datasets contain features related to wind speed, a correlation analysis was performed to validate the integration of the datasets (see
Figure 3). Features from the NORA3 reanalysis dataset have a high correlation factor of 0.85 with ground truth values obtained from LEGMC and nacelle sensors at corresponding heights.
The energy production profile of each wind turbine was recorded for the corresponding period (see
Figure 4). Most of the time, the output remains in the lower range of the turbine’s nominal capacity, which aligns with the wind rose characteristics described earlier. The dataset also includes intervals of zero production. While some of these periods correspond to planned maintenance activities, others may result from unexpected outages or adverse weather conditions.
During data preprocessing, intervals with zero energy production were excluded in cases where wind speed remained non-zero. This decision was made to prevent confusion in the learning process: under such conditions, the model might incorrectly learn that wind has no effect on energy production, which contradicts the operational behaviour observed during normal functioning.
Furthermore, wind speed data originated from multiple sources (e.g., nacelle sensors, meteorological stations, and reanalysis data), each exhibiting varying degrees of correlation with energy output. Since these sources are not always perfectly aligned, imputing or synthesising missing values could have introduced inconsistencies or biases into the dataset. Therefore, no artificial data generation or interpolation was applied in these cases to preserve the integrity of the original records.
Another site-specific feature of the energy production profile is its non-uniform dependency on wind direction at a given wind speed (see
Figure 5). The overall trend suggests that west-side winds yield higher power production compared to east-side winds. This can be explained by the laminar flow of air masses coming from the sea, which results in higher energy production.
Overall, the quality of the datasets is considered acceptable for the short-term forecasting task. Correlation analysis suggests consistency between various sources of wind parameters. Values of the parameters were within expected distributions without major outliers. Due to the origins of the dataset sources (provided by agencies specialising in long-term environment monitoring) the wind parameters formed a continuous timeline without missing values. In regards to the Nacelle Sensor and energy production values, the missing values were excluded from further processing.
3. Results
Machine learning provides practical methods for short-term wind speed forecasting, addressing the challenges of wind variability. By utilising historical data and advanced algorithms, these models enhance the accuracy of wind speed predictions and support wind farm operations.
To effectively train the model, it is necessary first to analyse and prepare the data. Initially, all datasets were merged into a single file, which allows for more convenient data analysis. The data format was standardised across all datasets.
The input features were normalised using the RobustScaler (scikit-learn version: 1.6.1), a scaling method based on the median and interquartile range, in order to reduce the impact of outliers. Rapid changes in wind speed, often caused by outliers that cannot be fully eliminated, negatively affect the dataset and reduce its overall stability.
Next, the records were filtered by time to exclude mismatched or incomplete time intervals, as well as synchronise records across datasets. As a result of the preprocessing procedures, the records are aggregated into a uniform hourly format. This temporal resolution serves as the foundation for the predictive modelling task in the current study, which aims to produce forecasts for a 36 h horizon, using the previous 36 time steps as input features. For training neural network models, the following hyperparameters were used: batch size = 64, number of epochs = 100, and learning rate = 0.001. Computations were executed on a CUDA-enabled GPU if available; otherwise, the CPU was used. For a more detailed overview of the dataset structure and features, refer to
Table 3.
The next step in data preparation involves analysing the correlation between the wind turbine’s produced energy and other parameters. Among the variables examined, wind speed, wind gusts, and wind direction exhibit the strongest correlations with energy output. While using these parameters together improves model results compared to analysing individual features, it also increases training time due to the higher dimensionality.
When evaluating wind speed data from different sources, NORA3 data demonstrates a higher correlation with observed energy production than LEGMC data. Further details of these comparative results are presented in
Figure 6.
To assess the influence of environmental factors on energy production, a correlation analysis was conducted using Pearson correlation coefficients. The results, shown in
Figure 6, highlight the degree of linear association between energy output (in kWh) and various meteorological and wind-related variables.
The analysis reveals that the wind speed measured by the Nacelle Sensor exhibits the strongest positive correlation with energy output (r = 0.95), indicating that it is a highly reliable predictor of actual energy generation. This is followed closely by wind speed received at 50 m from the NORA3 reanalysis dataset (r = 0.72) and wind speed at 6 m from the LEGMC station (r = 0.71). The wind gust at 10 m exhibits a moderate correlation (r = 0.31), whereas the wind direction shows a weaker relationship (r = 0.22).
Among meteorological variables, surface air pressure is negatively correlated with energy output (r = −0.31), suggesting that lower pressure conditions may be associated with higher wind activity and energy production. In contrast, temperature (r = −0.10) and relative humidity (r = 0.07) at 2 m show negligible correlation, implying minimal direct influence on energy generation in this context.
These results highlight the significant influence of wind speed measurements, particularly those recorded by onsite sensors, on determining the power output of wind energy systems.
During the analysis of the relationship between wind speed from the NORA3 dataset, LEGMC dataset, and energy production, it was found that the original data is characterised by a high level of noise and significant dispersion (
Figure 7). For identical wind speed values, there is a wide spread in the amount of generated energy, indicating the influence of random factors such as atmospheric turbulence, unstable weather conditions, short-term technical deviations, and other external influences.
To smooth the data and identify the underlying trend, a moving average method was applied, allowing for a reduction in variability without significantly distorting the original relationship. Due to the random nature of the noise, it is not feasible to make a precise prediction for each individual point. Therefore, the focus of the analysis shifts to determining the overall trend and value range.
The smoothed time series obtained through the moving average were subsequently used as input features for model training, which helped reduce the impact of short-term fluctuations and improve the model’s ability to generalise.
The models were trained using hourly data collected from 2022 through the end of 2024, which represents a relatively small dataset for time series forecasting (
Table 3).
Nevertheless, a total of eight models were developed and evaluated to assess their performance in predicting energy output. These included both traditional machine learning algorithms and deep learning architectures: linear regression, Ridge regression, XGBoost, CatBoost, LightGBM, Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Gated Recurrent Unit (GRU).
Table 4 below provides an overview of these models.
During the data preparation stage, it was necessary to format the data to be compatible with different model types. Since only recurrent (RNN) and convolutional (CNN) networks can process sequential data directly, other models such as fully connected neural networks (ANN), gradient boosting models (XGBoost, CatBoost, LightGBM), and statistical models (linear regression, Ridge) require a transformation. This conversion involved flattening all time steps and features into a single long vector, making the sequential structure compatible with non-sequential models.
Overall, the dataset comprised a limited number of time series, with a maximum of 12,000. The models showed a strong tendency towards overfitting. Consequently, the model parameters were adjusted to minimise the risk of excessive fitting to the training data. This ensured optimal generalisation capability for the models, guaranteeing their dependable performance on test data.
The use of error metrics in machine learning (ML) for energy forecasting is crucial for evaluating the accuracy and reliability of predictive models. The choice of evaluation metrics can significantly influence the interpretation of model performance and inform the refinement of forecasting techniques. In this study, the following commonly used error metrics were applied to assess model accuracy on the test data: Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-squared (R
2). When comparing model performance, the most favourable results correspond to lower values of MSE, MAE, RMSE, and MAPE, alongside a higher value of R
2. These metrics were used to evaluate all eight models across three distinct test datasets: Nacelle Sensor Testing Data, NORA3 Testing Data, and LEGMC Testing Data. The detailed results of model performance on each dataset are presented in
Table 5,
Table 6 and
Table 7, respectively.
Table 5 presents the results of eight models. When comparing models, it is essential to note that the best metrics are those with the lowest values of MSE, MAE, RMSE, and MAPE, as well as the highest value of R
2. As shown in the table, the linear regression model demonstrated the best values for MSE, RMSE, R, and MAPE. However, the XGBoost model achieved the lowest mean absolute error (MAE). It is worth noting that the metric values remain relatively high, which may be attributed to the limited amount of raw data used for training the models. This lack of data could have constrained the models’ ability to generalise and negatively impacted the quality of predictions.
Table 5.
Performance comparison of machine learning models on Nacelle Sensor testing data (Best results for each metric are bolded and underlined).
Table 5.
Performance comparison of machine learning models on Nacelle Sensor testing data (Best results for each metric are bolded and underlined).
Models | MSE | MAE | RMSE | R2 | MAPE |
---|
Linear Regression | 0.5975 | 0.6165 | 0.773 | 0.1391 | 2.739 |
Ridge | 0.5977 | 0.6166 | 0.7731 | 0.139 | 2.739 |
XGBoost | 0.5995 | 0.6100 | 0.7743 | 0.1362 | 1.981 |
CatBoost | 0.6015 | 0.616 | 0.7756 | 0.1334 | 2.241 |
LightGBM | 0.6568 | 0.6523 | 0.8104 | 0.0538 | 1.812 |
ANN | 0.6536 | 0.6418 | 0.8085 | 0.0583 | 2.128 |
CNN | 0.6558 | 0.6367 | 0.8098 | 0.0552 | 2.398 |
GRU | 0.6413 | 0.6262 | 0.8008 | 0.0761 | 2.143 |
Table 6 presents a comparison of the performance quality of several models on the NORA 3 dataset. All metrics exhibited the best values in linear regression, indicating their superiority in this task.
Table 6.
Performance comparison of machine learning models on NORA3 testing data (Best results for each metric are bolded and underlined).
Table 6.
Performance comparison of machine learning models on NORA3 testing data (Best results for each metric are bolded and underlined).
Models | MSE | MAE | RMSE | R2 | MAPE |
---|
Linear Regression | 0.6722 | 0.6325 | 0.8199 | 0.1684 | 1.488 |
Ridge | 0.6749 | 0.6339 | 0.8215 | 0.1651 | 1.489 |
XGBoost | 0.7274 | 0.6539 | 0.8529 | 0.1001 | 1.582 |
CatBoost | 0.7406 | 0.6655 | 0.8606 | 0.0838 | 1.527 |
LightGBM | 0.7802 | 0.6981 | 0.8833 | 0.0348 | 1.387 |
ANN | 0.7518 | 0.6658 | 0.8671 | 0.0699 | 1.295 |
CNN | 0.8288 | 0.6886 | 0.9104 | −0.0253 | 1.294 |
GRU | 0.7634 | 0.6676 | 0.8737 | 0.0556 | 1.229 |
Table 7 presents a comparison of the performance of various models on the LEGMC dataset. Linear regression achieved the most favourable results across all metrics, emphasising its effectiveness for this particular task. Based on the results shown in
Table 4,
Table 5 and
Table 6, it can be concluded that linear regression yields the highest prediction accuracy among the other models. This is attributed to the high degree of linear correlation in the data, which enables linear regression to approximate the dependencies between features and the target variable accurately. The Ridge regression model demonstrates comparable results, as it is essentially the same linear model but with L2 regularisation. Regularisation aids in combating overfitting, especially in the presence of multicollinearity, but renders the model less flexible and potentially more resource-intensive to train due to the additional parameter tuning required.
Table 7.
Performance comparison of machine learning models on LEGMC testing data (Best results for each metric are bolded and underlined).
Table 7.
Performance comparison of machine learning models on LEGMC testing data (Best results for each metric are bolded and underlined).
Models | MSE | MAE | RMSE | R2 | MAPE |
---|
Linear Regression | 0.7058 | 0.6685 | 0.8401 | 0.0654 | 1.386 |
Ridge | 0.7064 | 0.6689 | 0.8405 | 0.0646 | 1.389 |
XGBoost | 0.79 | 0.7068 | 0.8888 | −0.0462 | 1.548 |
CatBoost | 0.772 | 0.6972 | 0.8786 | −0.0223 | 1.628 |
LightGBM | 0.7982 | 0.7166 | 0.8934 | −0.0570 | 1.318 |
ANN | 0.7658 | 0.6941 | 0.8751 | −0.0140 | 1.459 |
CNN | 0.798 | 0.7032 | 0.8933 | −0.0566 | 1.302 |
GRU | 0.7589 | 0.684 | 0.8711 | −0.0050 | 1.349 |
Gradient boosting models and neural networks, in turn, possess significantly greater complexity and a larger number of parameters. This makes them more powerful when working with large, complex, and non-linear datasets. However, in cases where the data exhibits explicit linear dependencies, such models can overcomplicate the task and require more training time without a substantial gain in quality. Furthermore, due to their flexibility, they are more prone to overfitting on small datasets.
Figure 8 illustrates the performance of the linear regression model in predicting wind turbine energy output using three different datasets: Nacelle Sensor, NORA3, and LEGMC. The plot shows the actual wind turbine output (solid black line) compared with model predictions from each dataset over the time from late October to late November 2024.
The R-squared values in the range of 0.1–0.2, presented in
Table 5,
Table 6 and
Table 7, characterise the aggregated performance of the models over the entire 36 h forecasting horizon. This result is explained by the complexity of forecasting of highly stochastic processes in wind energy, where accuracy inevitably decreases with an increasing horizon. For example, for one of the models, the R-squared is approximately 0.95 for the 1st hour, 0.47 for the 18th hour, and 0.14 for the 36th hour, illustrating the typical behaviour of time series models.
Even moderate R-squared values at distant horizons indicate a significant reduction in uncertainty. High accuracy within the first 6 h ensures practical applicability for operational management, while forecasts for 24–36 h hold value for strategic planning.
The achieved indicators were also influenced by the limited volume of training data and the use of a moving average (window 5) to balance noise suppression and dynamics preservation. This method contributed to ensuring the stability of the model training.
Wind energy systems are characterised by an irreducible error caused by random factors, which limits the achievable accuracy. The goal of forecasting in this context is to reduce uncertainty to a level sufficient for making informed decisions, rather than achieving absolute accuracy. The presented results confirm the attainment of this goal.
The plot presents the linear regression predictions based on various datasets, alongside their corresponding Median Absolute Percentage Error (MdAPE) values, which indicate the percentage prediction error. Among the evaluated datasets, the Nacelle Sensor data produced the most accurate predictions, achieving a median absolute percentage error (MdAPE) of 8.72%. The NORA3 dataset yielded the most accurate predictions, with an MdAPE of 16.11%, while the LEGMC dataset produced the least accurate predictions, with an MdAPE of 21.85%. It is also evident that prediction accuracy varies over time. Specific time intervals display relatively accurate predictions across all three datasets, whereas projections in other periods are less reliable. This suggests that model performance is influenced by temporal variability and possibly by the quality or resolution of the input data during specific periods.
4. Conclusions
This study emphasises the crucial role of accurate short-term wind energy forecasting in ensuring reliable grid operation and effective participation in electricity markets. Enhanced forecast accuracy not only reduces balancing costs for energy providers but also enables greater integration of wind energy into power systems, an essential component of achieving long-term decarbonization goals. Using operational data from wind turbines in Latvia, supplemented by meteorological inputs from the NORA3 reanalysis dataset and LEGMC measurements, a range of machine learning and deep learning models was developed to generate energy output forecasts with lead times from 1 to 36 h.
The comparative evaluation of eight models, including linear regression, Ridge regression, XGBoost, CatBoost, LightGBM, ANN, CNN, and RNN, revealed that the quality and source of the input data significantly impact forecasting accuracy. Among the tested datasets, turbine-based operational data (Nacelle Sensor) yielded the most accurate results, followed by NORA3 reanalysis data, and finally LEGMC observations.
Linear regression emerged as the most effective model in this context due to the strong linear correlations present in the data and its computational efficiency. In contrast, complex models such as neural networks and gradient boosting did not yield substantial performance gains and were more prone to overfitting, particularly given the dataset’s limited size.
The superior performance of NORA3 over LEGMC data highlights the value of reanalysis datasets, which provide smoothed, noise-reduced meteorological inputs that are better aligned with modelling needs.
Additionally, the analysis of the relationships between input parameters and energy output revealed that wind speed measured directly at the turbine is the most reliable predictor of actual energy generation. Less accurate, but still useful predictors were obtained from reanalysis data and meteorological station measurements. While wind direction and gusts demonstrated some predictive value, they were less informative than wind speed. Environmental variables, such as temperature, humidity, and atmospheric pressure, exhibited a minor but noticeable impact on the predicted energy output.
Overall, the findings confirm that, when aligned with high-quality input data, even simple models can deliver decent short-term energy forecasts with a prediction horizon of up to 36 h using two years of historical data. At the same time, the simplicity of the model is its weakness—it cannot incorporate subtle nuances and random fluctuations of the wind parameter changes, which is suggested by relatively low metric values.
The study provides a proof-of-concept for the short-term wind energy production forecast problem. Further research is needed to generalise this approach for other locations and wind turbines.
Despite the promising results, this study has several limitations that should be acknowledged. Firstly, the observed relationship between wind speed and energy output is approximately linear. As a result, using complex model architectures such as Transformers, Informer, or N-BEATS is unlikely to deliver significant improvements. It may lead to overfitting, especially given the relatively low signal-to-noise ratio in the data. Secondly, the dataset exhibits high levels of noise, which complicates long-term forecasting. While data smoothing or aggregation techniques (e.g., hourly or daily averaging) can improve model stability, they often produce forecasts that capture general trends rather than exact values. Thirdly, the dataset is limited in scope both temporally and geographically. It covers only two years and is specific to a single turbine model at a fixed height and in a particular location. This raises questions about the wider applicability of the results to other turbine models, hub heights, terrains, and climate zones.