A Hybrid Deep Learning Framework for National Level Power Generation Forecasting of Different Energy Sources Including Renewable Energy and Fossil Fuel

Abstract

Electricity demand in the United States is steadily increasing due to rapid technological growth, especially the expansion of AI data centers and electric vehicles, which are becoming major power consumers. At the same time, rising renewable energy integration, changing weather patterns, and the deployment of battery energy storage systems are increasing variability and complexity in grid operations. These evolving conditions require advanced forecasting methods to ensure reliability and efficiency, as traditional statistical and machine learning models struggle with nonlinear and temporal dependencies. To address these challenges, this study proposes a hybrid deep learning framework that combines convolutional neural networks, long short-term memory, and bidirectional LSTM models to forecast electricity generation across both conventional and renewable energy sources. The framework incorporates seasonal-trend decomposition using loess to extract trend, seasonal, and residual components, enhancing the learning of multi-scale temporal patterns. A key contribution of this work is the development of a unified, source-specific forecasting system in which each energy source is assigned its best-performing hybrid architecture. The proposed framework achieves superior accuracy, with the CNN-Bi-LSTM model yielding the best total power results (MAPE 2.60%, RMSE 13,745 MWh, MAE 9542 MWh), while Bi-LSTM models excel for wind, biomass, geothermal, and nuclear. This enables scalable, high-precision national-level forecasting.

1. Introduction

The United States power sector is undergoing a major shift driven by technological innovations, policy mandates, and global environmental concerns. The power requirements are increasing steadily due to the advancement of technology and the growing application of AI in commercial and industrial sectors. According to the U.S. Energy Information Administration, the United States’ electricity consumption grew by 2% in 2024, and it is forecasted to continue growing at the rate of 1.7% in 2025 and 2026 [1]. To meet the demand, multiple power generation sources and technologies will be needed to generate electricity. The three main categories of energy sources for electricity generation are fossil fuels (natural gas, coal, and petroleum), nuclear energy, and renewable energy. From a technological perspective, steam turbines are responsible for generating the most electricity by utilizing fossil fuels, nuclear, biomass, geothermal, or solar thermal energy. The other electricity generation technologies are gas turbines, hydro turbines, wind turbines, and solar photovoltaic [2].

The large number of distributed energy resources (DERs), smart grid technologies, and the integration of renewable energy sources, particularly solar and wind, are redefining the operational landscape of electricity generation and distribution [3]. The high level of variability in the energy supply, especially from intermittent renewable sources, requires robust forecasting models to ensure grid reliability [4]. Accurate forecasting of power generation from different energy sources is critical for grid stability, optimizing reserve margins, minimizing operational cost, energy policy formulation, and infrastructure investment. Forecasting power generation helps in demand-side management, energy trading, and reducing carbon footprints by improving the integration of renewable sources into the grid [5,6].

In addition, accurate forecasts support the utility operators and system planners in balancing between supply and demand, scheduling generation units, and avoiding blackouts or power shortages [7]. Moreover, machine-learning-based predictive models play a crucial role in energy policy formulation by providing data-driven insights that inform investment decisions in infrastructure, storage, and transmission networks [8]. From a business perspective, precise electricity generation forecasting will help manage wholesale electricity markets effectively by reducing the capacity charges of the idle power plants and efficient energy trading [9].

Furthermore, from the perspective of sustainability, better forecasts contribute to lowering carbon emissions by facilitating the higher penetration of renewable energy sources and reducing fossil fuel-based plants [10]. For instance, integrating accurate renewable generation forecasts allows for the deployment of flexible backup systems such as battery storage or demand response mechanisms, thereby supporting environmental objectives [11].

Time series forecasting of electricity generation has mostly relied on statistical methods such as ARIMA, Holt–Winters exponential smoothing, and regression-based models [12]. While these models offer interpretability, they often fall short in capturing complex, non-linear patterns and long-term temporal dependencies present in real-world energy data [12]. Several forecasting methods, such as fuzzy neural networks [13], gray algorithms [14], gray Markov models [15], and support vector regression algorithms [16] have been proposed to forecast electricity generation. But these models are not efficient in managing large volumes or complex datasets [13,14,15,16]. In the field of photovoltaic (PV) forecasting, artificial neural networks (ANNs) and support vector machine (SVM) [17,18] are the most used machine learning algorithms. Several studies have shown that ANNs outperform traditional time series models when working with nonlinear datasets. However, ANNs often face challenges when handling large datasets and complex data mining processes [19].

To overcome these limitations, researchers have increasingly turned to deep learning-based machine models, particularly recurrent neural networks (RNNs) and their variants such as long short-term memory (LSTM) networks [20]. LSTM models have shown promise in energy forecasting due to their ability to learn long-term dependencies in sequential data [21]. However, standard LSTM networks process data in a unidirectional manner, which might not fully exploit the temporal relationships between past and future data points. Bidirectional LSTM (Bi-LSTM) networks address this by processing data in both forward and backward directions, providing a more comprehensive understanding of temporal dynamics [22].

1.1. Research Gap

Although hybrid deep learning methods have been widely studied for power and load forecasting, prior work is typically limited in one or more of four ways: (i) it focuses on a single source such as wind, solar, or load; (ii) it emphasizes short-term operational forecasting rather than long-horizon monthly generation patterns; (iii) it is restricted to regional or plant-level datasets rather than a coordinated national-scale setting; or (iv) it applies one model family uniformly rather than selecting source-specific architectures for heterogeneous generation types. In addition, several prior studies use decomposition-enhanced or hybrid architectures, but they do not formulate the forecasting problem as a unified source-wise model-selection framework across both renewable and fossil-fuel generation categories.

Motivated by this gap, this study develops a unified national-level monthly forecasting framework for U.S. power generation in which each source is assigned its best-performing deep architecture from a common candidate set. The novelty of the work lies not simply in applying CNN, LSTM, or Bi-LSTM models, but in integrating STL-based decomposition with source-specific model selection across heterogeneous generation sources under one consistent forecasting pipeline. This enables a coordinated yet source-adaptive framework for comparing and deploying forecasting models across coal, natural gas, petroleum-based, nuclear, hydro, wind, solar, geothermal, biomass, and aggregate total generation.

1.2. Core Contribution of This Research

The contributions of this work are:

• A novel application of a hybrid deep learning-based multi-model framework that combines CNN, LSTM, and Bi-LSTM for forecasting monthly electricity generation in the U.S. from multiple energy sources, including both conventional (e.g., coal, natural gas) and renewable (e.g., solar, wind, hydro) sources. This approach is designed to handle the diverse temporal patterns and nonlinearities present in multi-source power generation data.
• As part of the framework, six distinct hybrid models are developed and evaluated, each leveraging different configurations of CNN, LSTM, and Bi-LSTM layers to identify the optimal model structure for accurate forecasting across energy types.

To improve the quality of learning and enable the models to handle diverse seasonal patterns and trends in electricity generation, STL (seasonal-trend decomposition using loess) has been introduced as a data preprocessing step. This technique decomposes the original time series into three components: trend, seasonal, and residual, allowing the model to learn from each element independently.

2. Literature Review on Deep Learning: Forecasting Point of View

Neural networks, which form the foundational basis of deep learning, are particularly well-suited for analyzing monthly and quarterly time series, discontinuous series, and for generating forecasts extending several periods into the future. This literature review will systematically examine key deep learning architectures. The discussion will emphasize their theoretical and practical applications, and specific relevance to the complex task of power generation forecasting from both renewable and fossil fuel-based sources in the USA.

Table 1. Summary of deep learning architectures for time-series forecasting.

Architecture	Key Feature	Strength
CNN	Convolutional filters extract local features	Capture short term patterns
RNN	Sequential feedback structure	Temporal modeling
LSTM	Memory cell with gating mechanism	Long term dependency learning
Bi-LSTM	Bidirectional sequence processing	Improved context learning
CNN LSTM	CNN and LSTM Layers	Spatial and temporal learning
CNN-Bi-LSTM	CNN and Bidirectional learning	Strong nonlinear modeling

summarizes the deep-learning architectures considered in this study, highlighting their key features and typical strengths.

2.1. Convolutional Neural Networks (CNN)

CNNs has demonstrated remarkable utility and efficacy in time series forecasting. Their effectiveness in time series analysis comes from their ability to extract meaningful local patterns over time. This is achieved using one-dimensional (1D) convolutional kernels that move along the sequence, applying convolution operations to small, overlapping segments. Each segment is transformed into an embedding vector, often called a “token,” which captures important short-term features and trends within the time series [23]. A convolutional layer can be mathematically represented as:

$$y_t=\sigma \left(\sum _{i=0}^{k-1}\left(w_i·x_{t+i}+b\right)\right)$$

where y_t = output of convolutional operation, W_i = 1D convolutional filter (kernel) of size K, x_t+i = input time series of length T, $\sigma$ = activation function, b = bias of output map.

Figure 1 illustrates the conceptual architecture of a typical 1D CNN model for timer series forecasting, depicting the dataflow from an input layer receiving the 1D time series.

This input is fed into one or more 1D convolutional layers, where multiple filters are applied to extract various local features. Each convolutional layer is typically followed by an activation function and a pooling layer, which serve to reduce the dimensionality of the feature maps while preserving salient information. The outputs of these layers, representing the extracted local patterns, are then often flattened into a vector before being connected to subsequent layers, such as fully connected layers for direct prediction [24].

2.2. Recurrent Neural Networks (RNN)

Recurrent neural networks (RNNs) add feedback into a typical neural network. This allows the outputs to be based on immediate inputs as well as previous outputs of the network. This allows for greater pattern matching in time-series data.

A typical RNN is “unrolled,” and each complete neural network is referred to as a “cell”. Note that the biases for historical data in a basic RNN can introduce problems. When the bias is greater than 1, then the gradient explodes. When the bias is much less than 1, the gradient vanishes. This leads to a requirement for the bias to be close to, but not equal to 1, which may not be the best value for all systems. Extending the basic RNN model can solve this problem.

2.3. Long Short-Term Memory (LSTM)

Long short-term memory (LSTM) networks are an advanced form of RNNs specifically designed to learn long-term dependencies in sequential data. The main advantage of LSTMs is that they effectively address the vanishing gradient problem, which hampers learning over long sequences. The LSTM architecture includes a memory cell or cell state and three key gates. These are the input gate, forget gate, and output gate. The input gate controls which new information enters the memory cell, while the forget gate decides what information should be discarded from it. The output gate determines what information from the cell is passed on to the next time step. These gates work together to maintain and update the cell state over time. This mechanism enables LSTMs to maintain important context across long sequences, making them highly suitable for time series forecasting applications such as energy and power forecasting. The mathematical operations governing these gates and the cell state are as follows:

Forget gate (f_t): This gate determines which information from the previous cell state (c_t−1) should be discarded. It takes the current input (x_t) and the previous hidden state (h_t−1) as inputs and applies a sigmoid activation function (σ) to output values between 0 and 1. A value closer to 0 indicates forgetting, while a value closer to 1 indicates retention.

$$f_t=\sigma (W_f{x}_t+U_f{h}_{t-1}+b_f)$$

Input Gate (i_t): This gate decides which new information from the current input (x_t) and previous hidden state (h_t−1) should be stored in the current cell state. It also computes a candidate cell state (${\tilde{C}}_t$) using a hyperbolic tangent (tanh) activation function (σ), which scales values between −1 and 1.

$$i_t=\sigma (W_i{x}_t+U_i{h}_{t-1}+b_i)$$

$${\tilde{C}}_t=\sigma (W_c{x}_t+U_c{h}_{t-1}+b_c)$$

Cell State Update (c_t): The core memory of the LSTM is updated by combining the information retained from the previous cell state (modulated by the forget gate) and the new candidate information (modulated by the input gate). This element-wise multiplication and addition allow the cell state to carry relevant information forward across many time steps.

$$c_t=f_t{·c}_{t-1}+i_i·{\tilde{C}}_t$$

Output Gate (o_t): This gate controls which portions of the current cell state (c_t) are exposed and output as the hidden state (h_t). It uses a sigmoid function to filter the cell state, and the result is passed through a tanh function to produce the final hidden state.

$$o_t=\sigma (W_o{x}_t+U_o{h}_{t-1}+b_0)$$

$$h_t=o_t{tanh(c}_t)$$

where σ = activation function, f_t = forget gate, i_t = input gate, c_t = cell state update, o_t = output gate, ${\tilde{C}}_t$ = cell state, x_t = input vector at time t, h_t−1 = hidden state vector from the previous time step, c_t−1 = cell state vector from the previous time step, W and U = learned weight matrices, b = learned bias terms.

2.4. Bidirectional Long Short-Term Memory (Bi-LSTM)

Bidirectional long short-term memory (Bi-LSTM) networks extend conventional LSTM networks by processing an input sequence in both forward and reverse directions within the available input window. In this architecture, one LSTM layer reads the historical sequence from earliest to latest observation, while a second LSTM layer reads the same historical sequence in reverse order. The two hidden representations are then combined to produce a richer encoding of dependencies among the observations already contained in the input window. In the forecasting setting considered in this work, bidirectional processing is applied only to historical observations available at prediction time; it does not provide access to unseen future target values and therefore does not violate temporal causality.

Figure 2 illustrates the architecture used in this study. The model receives a historical sequence of time-series inputs and processes that sequence through LSTM or Bi-LSTM layers before passing the learned representation to a fully connected output layer. For Bi-LSTM, the backward branch operates only over the same historical input window, not over future observations beyond the prediction origin. Thus, the model benefits from richer within-window temporal encoding while preserving a causal forecasting setup.

In the case of Bi-LSTM, the model benefits from contextual information in both forward and backward directions, enhancing its ability to model complex temporal relationships. The features learned by the recurrent layers are subsequently passed to a fully connected layer, which integrates the temporal information and performs the final mapping to the output layer. This output layer generates predictions corresponding to the forecasted values. Overall, the architecture is particularly well-suited for time-series forecasting.

2.5. Deep Learning Based Hybrid Neural Networks

Deep learning-based hybrid neural network models have emerged as powerful tools for time series forecasting in the power and energy domain, owing to their ability to jointly capture spatial features and temporal dynamics. These models are typically constructed by integrating different neural network architectures, such as convolutional neural networks (CNNs), long short-term memory (LSTM), bidirectional LSTM (Bi-LSTM), gated recurrent units (GRUs), and attention mechanisms, to exploit their complementary strengths. For instance, CNN-LSTM and CNN-BiLSTM models are widely used in power forecasting, where CNN layers first extract high-level local features and patterns from input sequences, and the subsequent LSTM or Bi-LSTM layers model long-term dependencies across time steps.

In Section 3, a detailed exploration of previous works is provided. These works have successfully applied these hybrid models for short-term and long-term power forecasting, showing notable improvements over traditional statistical and single-model deep learning approaches. The integration of these architectures not only enhances prediction accuracy but also provides better generalization across varying time horizons and input feature complexities.

2.6. Evaluation Matrix for Forecasting Model

Mean absolute error (MAE) is a widely used statistical measure that quantifies the average magnitude of errors between predicted output values and their corresponding actual values within a dataset. It is computed by taking the sum of the absolute difference between each predicted value and its true value and then dividing by the total number of data samples [25]. The MAE is depicted as:

$$MAE={\displaystyle \frac{1}{N}}\sum _{i=1}^N|y_i-f_i|$$

where y_i is the actual value and f_i is the forecasted value for the power load and N is the number of data samples.

Root mean squared error (RMSE) is one of the most used measures for evaluating the quality of predictions in machine learning models. It is defined as the square root of the average squared difference in the actual value and prediction value, in other words, the square root of mean squared error (MSE) [25]. The RMSE is depicted as:

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}\sum _{i=1}^N{(y_i-f_i)}^2}$$

where y_i is the actual and f_i is the forecasted value for the power load and N is the number of data samples.

Mean absolute percentage error (MAPE) is a widely used metric for measuring the accuracy of forecasting models. It calculates the average of the absolute percentage errors between actual and predicted values. MAPE expresses the prediction error as a percentage, making it easier to interpret and compare across datasets with different scales. The MAPE is defined as:

$$MAPE={\displaystyle \frac{100}{N}}\sum _{i=1}^N|{\displaystyle \frac{y_i-f_i}{y_i}}|$$

where y_i is the actual value and f_i is the forecasted value for the power load, and N is the number of data samples.

2.7. Hybrid Deep Learning Models for Power Generation Forecasting

Over the past decades, various traditional and statistical methods have been developed for power generation forecasting and for planning the operation of energy systems [1,2,3,4,5,6,7,8,9,10]. However, increasing complexity and variability in power generation sources have led researchers to gradually shift focus toward more advanced machine learning techniques like deep learning models [11,12,13,14,15,16,17,18,19]. Hybrid deep learning models combine the strengths of multiple architectures like convolutional neural network (CNN), long short-term memory (LSTM), and other ML techniques that demonstrate superior performance in handling the non-stationary and highly dynamic nature of power generation data [20,21,22]. This review studies a range of hybrid deep learning models proposed for power generation forecasting using data from multiple energy/power generation sources, highlighting their methodologies, performance, and real-world applicability.

Table 2. Summary of previous hybrid deep learning forecasting studies.

Study	Model	Data Type	Forecast Horizon	Metrics
Gulay et al. [26]	STL-LSTM	Multi-source generation	Monthly	MAPE, RMSE
Ullah et al. [27]	CNN-M-BiLSTM	Residential load	Short-term	MSE, RMSE
Zafar et al. [28]	AE-LSTM	Solar generation	Short-term	MAE, RMSE
Zheng et al. [29]	CNN-A-LSTM-AR	Renewable multi-source	Short-term	MAPE
Shalini and Revathi et al. [30]	CNN-BiLSTM	Renewable hybrid system	Short-term	MSE, MAE
Rubasinghe et al. [31]	CNN-LSTM	Peak load	Long-term	MAPE

synthesizes prior hybrid deep-learning studies for power generation forecasting, including model classes, data types, forecast horizons, and evaluation metrics.

Recent studies have explored CNN-LSTM and CNN-BiLSTM architectures for renewable energy and electricity load forecasting. Several other works have reported improved short-term forecasting performance for wind and solar generation predictions and load predictions using CNN-LSTM structures. Similarly, CNN-BiLSTM models have enhanced bidirectional temporal learning capability in highly volatile renewable datasets. However, most of these studies focus on single-source forecasting, such as wind only, solar only, or load only prediction and short-term operational application rather than integrated multi-source power generation modeling.

A comprehensive analysis of electricity generation forecasting in Turkiye was completed for different energy sources such as natural gas, hydro, coal, wind, fuel oil, bioenergy, and geothermal power [26]. The study evaluated different types of hybrid models with the combination of decomposition, machine learning models (especially deep learning), and statistical models to predict the electricity production. According to the authors, to extract the meaningful components from non-linear time series data like seasonal trend and residual trends, STL decomposition (seasonal and trend decomposition using loess) has been leveraged with the models. The monthly periodical power generation data set from January 2012 to May 2022 was used to train and test those models. The hybrid deep learning model demonstrated the best forecasting performance across the entire variable. For evaluating and comparing the model, the authors used the mean absolute percentage error (MAPE) and root mean square error (RMSE) loss functions. Among the hybrid deep learning models, STL-LSTM (long short-term memory) verified optimal performance for four variables, and STL-CNN showed optimal performance for three variables. During the performance evaluation process, the hybrid model was also compared with eleven other models based on RMSE and MAPE.

A hybrid model via convolutional neural network (CNN) with multi-layer bidirectional long short- term memory (CNN-M-BDLSTM) was proposed for short-term residential power consumption prediction [27]. The household power consumption dataset—including date, time, global active power, reactive power, voltage, amperes, and sub-metering values from smart meters monthly—was used as the input dataset. The network of the proposed model consists of sequential convolutional layers, pooling layers, M-BDLSTM layers, and dense layers. To verify the effectiveness of the proposed model, it was evaluated with a holdout cross-validation method and 10-fold cross-validation method using various metrics, including mean squared error (MSE), RMSE, and mean absolute error (MAE). Additionally, a CNN-M-BDLSTM model was also compared with three other models named LSTM, BDLSTM, and CNN-LSTM. The proposed method outperforms these methods by achieving the smallest value for the MSE.

A hybrid forecasting model combining an auto encoder with long short-term memory (AE-LSTM) was proposed to improve the accuracy of short-term solar power generation prediction [28]. The model addresses the nonlinear and temporal nature of solar energy data by first compressing and denoising the input features using an AE, followed by temporal sequence learning through the LSTM. The dataset consists of daily power generation (kWh), maximum grid-connected power generation (MW), and radiation (MJ·m⁻²) over a year of duration. Experimental results demonstrate that the AE-LSTM model outperforms conventional LSTM and Bi-LSTM models in terms of accuracy and performance measures for each of the parameters. For evaluating and comparing the model, the authors used the MAE, MSE, and a RMSE matrix.

In [29], a hybrid framework model was proposed for forecasting renewable energy sources power generation that integrates with convolutional neural networks (CNN), attention-based long short-term memory (A-LSTM), and auto regressive (AR). The sources of prediction were solar PV, solar thermal, and wind energy in renewable energy systems (RES). The dataset consisted of power generation data from different renewable energy sources and weather data. This hybrid framework captured energy correlation patterns, nonlinear temporal dependencies from weather and generation data, and linear trends of each energy source. A comparison study has been conducted between the proposed hybrid model with A-CNN-LSTM, ANN, and eXtreme Gradient Boosting (XGB). The proposed method achieved superior accuracy by 9.16%, 18.15%, and 16.87% of MAPE for solar PV, solar thermal, and wind power respectively. Sensitivity analyses confirmed that incorporating energy correlations and nonlinear temporal patterns significantly enhanced forecasting performance.

A hybrid CNN-based Bi-LSTM method has been utilized for forecasting renewable energy power generation such as solar PV and wind by combining convolutional neural networks (CNN) and bidirectional long short-term memory (Bi-LSTM) [30]. The study also integrates a grid-connected hybrid energy system utilizing solar PV, wind turbines, a battery storage unit, a modified Z-source converter, and bidirectional DC–DC converters to ensure a stable DC link voltage. The CNN-Bi-LSTM model enhances forecasting by capturing both spatial and temporal patterns in renewable energy data, outperforming conventional models like CNN, LSTM, and Bi-LSTM in terms of mean squared error (MSE), mean absolute error (MAE), and R² accuracy metrics.

A comprehensive comparative analysis of various time series forecasting methods was conducted for predicting solar power generation from large-scale photovoltaic (PV) plants, particularly when direct solar irradiance or detailed weather data are unavailable. The authors evaluate statistical models like autoregressive moving average (ARMA), autoregressive integrated moving average (ARIMA), and seasonal autoregressive integrated moving average model (SARIMA) against several artificial intelligence-based approaches, including feedforward neural networks, multilayer perceptions (MLP), recurrent neural networks (RNN), LSTM, and a proposed Bi-LSTM model [32], using hourly operational data from a 20 MW PV plant in China. Key findings indicate that neural network models, especially Bi-LSTM, offer superior accuracy (R value of 0.98 for one-hour ahead predictions) and generally require less computation time compared to the statistical models for short-term forecasting.

A hybrid CNN-Bi-GRU [33] model was established for short and long-term renewable electricity demand forecasting, utilizing a multimodal information fusion approach that integrates time series data (like historical power generation) and textual data (such as policy documents). The methodology involves preprocessing both data types, then using CNNs to extract spatial features, which are subsequently fused via concatenation. These fused features are fed into a bidirectional gated recurrent unit (Bi-GRU) network to capture temporal dependencies, forming the complete CNN-Bi-GRU model. Comparative experiments against ARIMA, standalone GRU, and ensemble empirical mode decomposition (EEMD)-ARIMA models demonstrate the proposed model’s superior predictive performance in both short-term (monthly) and long-term (semi-annual) forecasting tasks on the same dataset, as evidenced by lower RMSE and MAPE values.

A novel sequence-to-sequence hybrid CNN-LSTM model [31] was developed for long-term monthly peak load forecasting, specifically for a three-year horizon. Addressing the drawbacks of classical methods like ANNs and regression models, such as insensitivity to long-term trends and difficulty with numerous variables, this work focused on monthly peak load to balance detail and computational efficiency for long-term tactical planning. The model verified using load data from New South Wales, Australia, incorporated data preprocessing (including PCHIP interpolation for annual data and Pearson correlation for feature selection), PCA for dimensionality reduction, and a sequence-to-sequence architecture where CNN acts as an encoder and LSTM as a decoder. The research demonstrates higher prediction accuracy compared to existing methods for long-term forecasting, with the CNN-LSTM model achieving a mean absolute percentage error (MAPE) of 4.29% for a 36-month ahead forecast, outperforming the standalone CNN (4.55% MAPE) and LSTM (4.91% MAPE) models. Additionally, the paper developed a block bootstrapping-based probabilistic forecasting framework to account for uncertainties in economic, population, and climate data, highlighting the model’s applicability for both point and probabilistic long-term tactical planning.

Taken together, the reviewed studies show that hybrid and decomposition-enhanced deep learning models can improve forecasting performance when temporal structure is nonlinear, multi-scale, or highly variable. However, most prior work remains targeted to individual sources, short-term horizons, or localized datasets, and only limited attention has been given to coordinated source-wise forecasting at the national scale. Moreover, although STL-based and hybrid architectures have been reported previously, the literature still lacks a framework that evaluates multiple deep architectures under one common monthly U.S. dataset and then assigns the most suitable model to each generation source. This unresolved gap motivates the present study.

3. Materials and Methods

The study aims to develop a complete framework for performing the time series forecasting of USA’s power generation using hybrid deep learning models. The dataset consists of electricity power generation data from different sources, such as conventional and renewable energy sources. In this section, the detailed process of collecting and processing power generation data has been discussed.

3.1. Data Collection and Data Split

The historical dataset obtained from the U.S. Energy Information Administration (EIA), which includes detailed electricity generation data in Figure 3, is categorized by sources such as coal, petroleum liquid, petroleum coke, natural gas, other gases, nuclear, conventional hydroelectric, wind, geothermal, biomass, all solar (small scale and utility scale), and other sources. Datasets are arranged in monthly periods and contain 159 observations from January 2012 to May 2025 [25]. These datasets provide publicly accessible, validated national-level electricity generation statistics categorized by fuel type.

The time series analysis of monthly power generation data reveals significant shifts in the energy mix of the United States. Coal-fired power generation has shown a marked and steady decline over the years, indicating a move away from this carbon-intensive source. Similarly, petroleum-based sources, such as liquid petroleum and coke petroleum, display either a downward trend or highly sporadic usage, suggesting their limited and declining role in the overall energy portfolio. On the other hand, natural gas has demonstrated a strong upward trajectory, emerging as the dominant fossil fuel, likely due to its economic and environmental advantages over coal. Renewable energy sources show a clear growth trend, particularly solar and wind. Solar energy grew rapidly from negligible levels in 2012 to become a substantial contributor by 2024, while wind energy has seen consistent and robust increases throughout the period. Geothermal power has remained relatively stable with minor fluctuations, and hydroelectric power, while seasonally variable, has not shown a long-term growth trend. Nuclear power generation has remained largely constant, with periodic seasonal fluctuations but no significant upward or downward shift. Meanwhile, biomass and other miscellaneous sources initially maintained stable output but began to decline slightly in recent years. Overall, the total power generation has increased gradually, with a noticeable transition from coal-dominated generation to a mix increasingly reliant on natural gas and renewables, particularly wind and solar.

Each power generation source is modeled using its own optimized deep learning architecture. These source-specific models operate simultaneously within a unified hybrid forecasting framework, allowing coordinated multi-source prediction without enforcing a single multivariate structure. The dataset is split chronologically into 80% training and 20% testing sets to preserve temporal order. From the training portion, 10% is further reserved for validation during model training. All preprocessing steps, including normalization, are fitted only on the training dataset. The learned parameters are then applied to validation and testing sets to avoid data leakage. Despite the limited number of monthly observations per source, the adopted forecasting models are intentionally capacity-constrained and trained under regularized, leakage-controlled conditions appropriate for a small-data regime.

3.2. Data Preprocessing by Decomposition Technique

To enhance the quality of forecasting and enable the model to learn distinct temporal patterns more effectively, we apply a decomposition-based data preprocessing approach prior to model training. Specifically, we use the seasonal-trend decomposition using loess (STL) technique to disaggregate the original time series into three fundamental components: trend, seasonal, and residual. This decomposition enables the separation of systematic patterns in the data, thereby simplifying the complexity of the forecasting task and improving model interpretability. Mathematically, the original time series Y(t) is expressed as the sum of its components:

$$Y\left(t\right)=T\left(t\right)+S\left(t\right)+R\left(t\right)$$

where T(t) represents the trend component, S(t) denotes the seasonal component, and R(t) is the residual component.

After decomposition, each component T(t), S(t), and R(t) is individually treated as a separate input feature to the deep learning architecture. This enables the model to learn unique temporal dependencies from each component without interference from unrelated patterns. For instance, the trend component contributes to long-term memory learning, while the seasonal part provides information about periodic patterns, and the residual highlights anomalies or short-term volatility. By integrating STL decomposition into the data preprocessing pipeline, the model benefits from a more structured and noise-reduced input space, ultimately leading to improved forecasting accuracy and generalization performance. This approach has been shown to be particularly effective in energy and load forecasting applications, where seasonal cycles and non-linear trends are prominent.

STL is applied separately to each of the power generation sources. Since the data is monthly, the seasonal period is set to 12 months. After decomposition, the trend, seasonal, and residual components are treated as separate input features to the deep learning model. Missing values are handled using linear interpolation followed by forward/backward filling if necessary. Outliers are mitigated through STL’s robust fitting mechanism. Min-max scaling to the range [0, 1] is applied. Scaling parameters are computed using only the training dataset and applied consistently to validation and test sets.

The electricity generation data used in this research exhibit strong seasonal and long-term structural patterns due to factors such as demand cycles, policy shifts, renewable energy growth. STL is particularly suitable for such datasets because it allows the seasonal component to vary over time and is robust to outliers and non-stationary behavior. By separating these components prior to model training, the learning algorithms can focus on capturing meaningful temporal patterns rather than simultaneously modeling trend, seasonality, and noise. This approach reduces the complexity of the learning task and improves interpretability of the forecasting framework. For these reasons, STL decomposition is incorporated as a preprocessing step within the proposed forecasting pipeline.

Although each source contains only 159 monthly observations, the forecasting models used in this study were intentionally capacity-constrained and regularized to suit a small-data setting. First, the STL preprocessing step transforms each series into three structured components (trend, seasonal, and residual), reducing the burden on the network to learn all sources of variation directly from the raw signal. Second, model training is performed on rolling historical windows rather than on a single sequence, which increases the number of supervised training instances while preserving chronological order. Third, the candidate architectures were restricted to comparatively small search ranges for filters, recurrent units, dense units, dropout, and learning rate, and model selection was performed using a validation set with early stopping to limit overfitting. A comparable monthly data scale has also been used in prior work which employed monthly power-generation data from January 2012 to May 2022 in an STL-based hybrid forecasting framework, supporting the relevance of decomposition-assisted learning in limited-sample monthly forecasting settings [26]. Accordingly, the objective of the proposed framework is not to deploy large-capacity deep networks, but rather to evaluate compact sequence models that can capture nonlinear and source-specific temporal structure under strict temporal-causality and regularization constraints. Nevertheless, the limited sample size remains a study limitation, and the reported findings should be interpreted as evidence of relative model suitability within this constrained setting rather than as a universal claim that deep learning is always preferable to simpler forecasting methods.

3.3. Proposed Power Generation Forecasting Framework by a Hybrid Deep Learning Model

In this section, a hybrid deep learning-based forecasting framework is presented to improve the accuracy of power generation prediction across various energy sources in the USA. The framework is structured in two main phases: first, identifying the best-performing model for each power generation source through a systematic evaluation of six deep learning architectures, and second, applying the selected models to build a unified, source-specific forecasting framework. This approach allows for tailored forecasting that accounts for the unique characteristics and patterns of different energy sources, both conventional and renewable.

3.4. Methodology of Selecting Source-Wise Best-Fitted Model

Figure 4 outlines the methodological flowchart employed to forecast power generation in the USA based on different energy sources. The process begins with the collection of power generation data. Following this, a data preprocessing step is performed using decomposition techniques to extract key components such as trend, seasonality, and residuals, which help improve the learning capability of the models. Subsequently, six distinct deep learning models are designed: CNN, LSTM, Bi-LSTM, CNN-LSTM, CNN-BiLSTM, and a stacked model combining CNN, Bi-LSTM, and CNN layers.

These models are then trained and evaluated using multiple performance metrics, including mean absolute percentage error (MAPE), mean squared error (MSE), and root mean squared error (RMSE). If the evaluation indicates the need for improvement, parameter tuning, and optimization techniques are applied iteratively. Once the models achieve satisfactory performance, the best-performing model is selected individually for each power generation source, enabling source-specific forecasting with higher accuracy.

To optimize model performance, hyperparameter tuning is conducted to identify the best parameter configurations for each architecture. A total of 30 trials were executed per model using Amazon SageMaker AI Studio’s online environment, which provided scalable cloud-based resources for efficient model training and evaluation. This process ensured that each model is fine-tuned to achieve its maximum forecasting accuracy before inclusion in the comparative analysis. The hyperparameter search space has been defined for all deep learning architectures for consistent comparisons in

Table 3. Hyperparameter search space for deep learning architectures.

Hyperparameter	Layer Description	Search Range	Step Size
cnn_filters1/filters	First CNN layer filters	32–128	32
cnn_filters2	Second CNN layer filters	16–64	16
kernel_size1	Kernel size (first CNN)	{2, 3, 4}	—
kernel_size2	Kernel size (second CNN)	{1, 2, 3}	—
cnn_dropout1	Dropout after first CNN	0.1–0.5	0.1
cnn_dropout2	Dropout after second CNN	0.1–0.5	0.1
lstm_units1	First LSTM/BiLSTM layer units	32–128	32
lstm_units2	Second LSTM/BiLSTM layer units	16–64	16
dropout1	Dropout after first LSTM	0.1–0.5	0.1
dropout2	Dropout after second LSTM	0.1–0.5	0.1
dense_units	Fully connected layer units	32–128	32
learning_rate	Adam optimizer learning rate	{0.001, 0.0005, 0.0001}	—

.

To ensure reproducibility and prevent information leakage, all preprocessing procedures, including missing-value handling, STL decomposition, and min-max scaling, were fitted exclusively on the training data and subsequently applied to the validation and test sets. Furthermore, in the Bi-LSTM models, bidirectional processing was confined to the ordered observations within each historical input window and did not involve unseen future target values. For each power generation source, forecasting was formulated as a supervised sliding-window time-series learning task, where historical observations were transformed into input-output sequences. In particular, the preceding 12 months of STL-derived features, namely the trend, seasonal, and residual components, were used to predict power generation one month ahead. Sliding windows were generated sequentially within each dataset split to ensure that no training sample contained information from the validation or test periods. All models were trained using the Adam optimizer, with learning rates selected from the hyperparameter search space presented in Table 3. Training was performed for a maximum of 100 epochs using a batch size of 32.

3.5. Proposed Framework by Using the Best-Fitted Model for Each Power Generation Source

The flow diagram depicted in Figure 5 shows the comprehensive and structured framework for forecasting power generation using a hybrid deep learning approach designed to handle the complexities of diverse energy sources. The process begins with:

• Step 1: Data collection—raw time series power generation data is gathered from various sources, including both conventional (such as coal, natural gas, nuclear, and petroleum-based fuels) as well as renewable sources (like wind, solar, geothermal, and biomass). This wide-ranging data provides a strong foundation for accurate forecasting by capturing the heterogeneous nature of the U.S. power grid.
• Step 2: Data preprocessing—the collected raw data undergoes decomposition using techniques like STL (Seasonal-Trend decomposition using loess), which separates each time series into three fundamental components: trend, seasonal, and residual. This decomposition enhances model learning by isolating long-term growth patterns, short-term seasonal cycles, and noise, allowing more focused training for each aspect of the data.
• Step 3: Model selection and training—designing and evaluating six hybrid deep learning models mentioned in Section V to determine the most effective model for each power generation source. These models are trained on the decomposed data and assessed using evaluation metrics like mean absolute percentage error (MAPE), mean absolute error (MAE), and mean squared error (MSE). The best-performing model for each energy source is selected through an iterative process involving training, performance evaluation, and hyperparameter tuning.
• Step 4: The proposed hybrid deep learning forecasting framework is established by integrating these source-specific models into a multi-model forecasting structure. In this stage, the decomposed data is loaded into the best-fitted model for each source, and the multi-model framework is trained and evaluated holistically.
• Step 5: Forecasting—the trained framework is used to forecast power generation for each energy source. The output forecasts are evaluated again using MAPE, MAE, and MSE to ensure the reliability and accuracy of the model. This step validates the effectiveness of the overall forecasting approach and demonstrates the benefits of using specialized models for different sources within a unified framework.

Overall, the proposed flow represents a robust, scalable, and intelligent system for power generation forecasting, with the flexibility to support grid modernization, policymaking, and the integration of renewable energy sources.

4. Simulation Results and Discussion

The performance of the proposed approach is evaluated using historical power-generation data across multiple energy sources. The evaluation uses common error metrics such as MAPE, RMSE, and MAE to quantify forecasting accuracy under a consistent experimental protocol. The comparative analysis in this study is focused on six deep-learning architectures evaluated under the same data partitioning and preprocessing pipeline, providing a consistent basis for selecting the most suitable source-specific hybrid model. A limitation of the present study is that the experimental comparison is restricted to the evaluated deep learning architectures and does not include direct benchmark tests against standard conventional baselines such as seasonal naïve forecasting, ARIMA/SARIMA, Prophet, or exponential smoothing using the same dataset partition. Therefore, the present results establish which evaluated neural architecture performs best within the proposed source-specific framework, but they do not yet establish whether the framework is superior to simpler and more interpretable forecasting methods. This benchmark comparison is an important next step and will be addressed in future work using the same chronological split, preprocessing pipeline, and evaluation metrics to enable a fair comparison.

The forecasting performance of six deep learning architectures CNN, LSTM, Bi-LSTM, CNN-LSTM, CNN Bi-LSTM, and CNN Bi-LSTM CNN has been evaluated across different energy generation sources using three complementary error metrics: mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE), which are depicted in

Table 4. Comparisons of forecasting performance in terms of MAPE (%).

Model	Total Power	Coal	Liq. Petro.	Coke Petro.	Natural Gas	Other Gases	Nuclear	Conv. Hydro	Wind	Geo-Thermal	Biomass	Solar	Other Sources
CNN	4.17	16.35	18.75	42.70	3.21	8.99	2.03	7.36	7.27	6.81	3.54	7.45	4.77
LSTM	3.79	14.95	19.97	24.28	4.48	10.13	2.78	7.78	8.34	4.88	3.75	12.99	16.01
Bi-LSTM	2.93	15.03	16.22	25.68	4.15	9.24	1.66	9.60	6.62	4.07	2.36	6.22	6.81
CNN-LSTM	3.52	16.84	21.68	91.85	4.69	10.90	1.70	7.33	9.06	4.97	3.45	20.98	13.27
CNN Bi-LSTM	2.60	22.12	16.38	86.64	4.21	7.39	1.89	8.01	6.89	6.21	5.33	12.90	16.05
CNN Bi-LSTM CNN	3.17	25.33	15.39	69.23	4.62	9.98	1.74	9.01	6.46	5.73	13.02	12.06	16.05

,

Table 5. Comparisons of forecasting performance in terms of RMSE.

Model	Total Power	Coal	Liq. Petro.	Coke Petro.	Natural Gas	Other Gases	Nuclear	Conv. Hydro	Wind	Geo-Thermal	Biomass	Solar	Other Sources
CNN	17,643	10,458	518	140	6033	95	1557	1940	3382	109	180	1893	43
LSTM	17,508	8660	544	114	10,274	108	2175	1928	3847	77	177	4224	132
Bi-LSTM	13,015	9233	522	108	7737	99	1217	2140	3110	70	111	1836	61
CNN-LSTM	17,441	9898	521	279	10,119	110	1283	1871	4257	78	181	6007	112
CNN Bi-LSTM	13,745	12,825	508	265	8606	82	1499	2141	3174	109	241	3417	136
CNN Bi-LSTM CNN	13,973	14,169	542	212	8762	111	1373	2239	3277	105	532	3392	133

, and

Table 6. Comparisons of forecasting performance in terms of MAE.

Model	Total Power	Coal	Liq. Petro.	Coke Petro.	Natural Gas	Other Gases	Nuclear	Conv. Hydro	Wind	Geo-Thermal	Biomass	Solar	Other Sources
CNN	14,391	8234	260	123	4791	76	1273	1427	2594	93	134	1550	35
LSTM	13,940	7508	282	88	6523	85	1758	1494	2933	66	142	3217	114
Bi-LSTM	10,795	7543	240	85	6151	75	1069	1791	2478	56	89	1368	50
CNN-LSTM	13,101	8304	285	249	6828	89	1106	1425	3520	67	129	5039	95
CNN Bi-LSTM	9542	11,049	233	239	6452	63	1201	1604	2426	85	202	2882	113
CNN Bi-LSTM CNN	11,262	12,501	238	190	6838	80	1128	1699	2502	80	502	2826	115

, respectively. MAPE provides a scale-independent measure of relative error, while RMSE and MAE, expressed in megawatt-hours (MWh), assess absolute error magnitude, with RMSE penalizing larger deviations more heavily.

Across the three metrics, no single model consistently outperformed all others for every energy source, highlighting the variability in source-specific generation patterns and the differing strengths of each architecture. However, hybrid models, particularly CNN-Bi-LSTM and Bi-LSTM, demonstrated superior accuracy in multiple categories, suggesting that combining convolutional feature extraction with bidirectional sequence modeling provides an advantage in capturing both local and long-term temporal dependencies.

Total power: For total power generation, CNN Bi-LSTM achieved the best results in all three metrics (MAPE: 2.60%, RMSE: 13,745 MWh, MAE: 9542 MWh), confirming its robustness for large-scale forecasting. This indicates that hybrid architectures are well-suited for integrating the diverse variability of multiple generation sources.

Fossil fuels: Coal generation was best predicted by LSTM in terms of MAPE (14.95%) and RMSE (8660 MWh), while its MAE was also among the lowest (7508 MWh), highlighting the importance of sequential modeling for this source. In liquid petroleum and coke petroleum forecasting, Bi-LSTM achieved the lowest errors across all metrics, confirming the benefit of bidirectional temporal modeling for volatile, smaller-scale generation patterns. For natural gas, CNN achieved the lowest RMSE (6033 MWh) and MAPE (3.21%), whereas Bi-LSTM had the lowest MAE (6151 MWh), suggesting CNN’s better handling of extreme deviations and Bi-LSTM’s advantage in reducing average errors.

Gaseous sources: In other gases, CNN Bi-LSTM consistently outperformed other models (MAPE: 7.39%, RMSE: 82 MWh, MAE: 63 MWh), reflecting the effectiveness of convolutional feature extraction followed by bidirectional sequence learning for small but variable sources.

Nuclear and hydro power: Nuclear generation forecasting was most accurate with Bi-LSTM (MAPE: 1.66%, RMSE: 1217 MWh, MAE: 1069 MWh), while conventional hydro performance peaked with CNN Bi-LSTM CNN in RMSE (1373 MWh) and MAE (1128 MWh), and CNN-LSTM for MAPE (7.33%).

Renewables: Wind generation, characterized by high intermittency, was best predicted by Bi-LSTM (MAPE: 6.62%, RMSE: 3110 MWh, MAE: 2478 MWh). For geothermal, Bi-LSTM again led with the lowest RMSE (70 MWh) and MAE (56 MWh), as well as the best MAPE (4.07%). Biomass forecasting showed the smallest errors with Bi-LSTM (MAPE: 2.36%, RMSE: 111 MWh, MAE: 89 MWh). Interestingly, solar generation forecasting favored Bi-LSTM in RMSE (1836 MWh) and MAE (1368 MWh), although CNN achieved the lowest MAPE (7.45%).

Miscellaneous sources: For other sources, CNN consistently achieved the lowest errors (MAPE: 4.77%, RMSE: 43 MWh, MAE: 35 MWh), indicating its strength in stable, small-scale sources.

The integrated evaluation shows that Bi-LSTM dominated for most individual energy sources, particularly those with high variability or smaller generation scales, while CNN Bi-LSTM emerged as the most reliable all-around model for aggregated and mixed-source forecasting. The results confirm that architectures combining convolutional and bidirectional recurrent layers offer a balanced capability to extract spatial features, capture temporal dependencies, and handle both average and deviations in multi-source energy forecasting.

4.1. Analysis of Proposed Hybrid Deep Learning Forecasting Framework

The comparative evaluation across MAPE, RMSE, and MAE metrics demonstrates that no single deep learning architecture consistently delivers the best performance for all energy sources. Each source exhibits distinct temporal dynamics, seasonal variability, and volatility patterns, meaning that certain architectures capture its characteristics more effectively than others. To address this, the proposed framework adopts a multi-model forecasting strategy in which the best-performing model for each source identified through the evaluation in Table 4, Table 5 and Table 6, is selected for its respective prediction task (

Table 7. Best performing model selection for the proposed forecasting framework.

Energy Source	Selected Model	Reason for Selection (Metric Reference)
Total Power	CNN Bi-LSTM	Lowest MAPE (2.60%), RMSE (13,745 MWh), and MAE (9542 MWh)
Coal	LSTM	Lowest MAPE (14.95%) and RMSE (8660 MWh); competitive MAE
Liquid Petroleum	Bi-LSTM	Lowest RMSE (522 MWh) and MAE (240 MWh); competitive MAPE
Coke Petroleum	Bi-LSTM	Lowest RMSE (108 MWh) and MAE (85 MWh)
Natural Gas	CNN	Lowest MAPE (3.21%) and RMSE (6033 MWh)
Other Gases	CNN Bi-LSTM	Lowest MAPE (7.39%), RMSE (82 MWh), and MAE (63 MWh)
Nuclear	Bi-LSTM	Lowest MAPE (1.66%), RMSE (1217 MWh), and MAE (1069 MWh)
Conventional Hydro	CNN Bi-LSTM CNN	Lowest RMSE (1373 MWh) and MAE (1128 MWh)
Wind	Bi-LSTM	Lowest MAPE (6.62%), RMSE (3110 MWh), and MAE (2478 MWh)
Geothermal	Bi-LSTM	Lowest MAPE (4.07%), RMSE (70 MWh), and MAE (56 MWh)
Biomass	Bi-LSTM	Lowest MAPE (2.36%), RMSE (111 MWh), and MAE (89 MWh)
Solar	Bi-LSTM	Lowest RMSE (1836 MWh) and MAE (1368 MWh)
Other Sources	CNN	Lowest MAPE (4.77%), RMSE (43 MWh), and MAE (35 MWh)

). For instance, Bi-LSTM is employed for wind, biomass, and geothermal due to its superior handling of long-term dependencies and high variability, while CNN Bi-LSTM is selected for total power and other gases for its ability to integrate spatial feature extraction with bidirectional temporal learning. Similarly, LSTM is chosen for coal, CNN for natural gas and other sources, and CNN Bi-LSTM CNN for conventional hydro.

Once the optimal model is assigned to each source, forecasts are generated independently and then aggregated to obtain the total power generation prediction. This modular approach ensures that the unique advantages of different architectures are fully leveraged, resulting in improved overall accuracy and robustness compared to a single-model approach. Figure 6 and Figure 7 present the predicted versus actual generation curves for each source using its respective best model. The close alignment of these curves across diverse generation types confirms that the proposed framework effectively captures both short-term fluctuations and long-term trends in multi-source power generation.

The comparative plots of actual and predicted generation in Figure 6 and Figure 7 across diverse energy sources highlight the capacity of the proposed multi-model framework to capture both seasonal patterns and source-specific fluctuations. The close alignment of predicted and observed curves for major sources such as natural gas, coal, solar, and total power suggests that the framework can reliably forecast energy generation trends, enabling policymakers and grid operators to anticipate supply levels with higher confidence. For renewable sources such as wind, solar, and hydro, accurate forecasting supports informed policy decisions regarding storage capacity planning, integration of intermittent sources, and investment prioritization. Similarly, for fossil fuel-based sources, these predictions can aid in developing transition strategies, emissions control measures, and contingency planning during demand peaks or supply disruptions. By providing a robust and timely view of future generation scenarios, the framework offers a valuable tool for evidence-based policymaking in the evolving energy sector.

4.2. Comparative Analysis with Previous Works

A comprehensive survey of deep learning-based hybrid models for forecasting in power systems has been derived in Section VIII as shown in

Table 8. Comparisons of proposed model with other previous hybrid deep learning models.

Model/Reference	Key Feature	Input Data Type	Application Domain	Reported Performance
Different Deep Learning Model [26]	Data decomposition before employing the deep learning and machine learning model	Monthly periodical power generation data and economic data	Forecasting of electricity generation of Turkiye	STL-LSTM performed best in terms of MAPE ranging 2.087–50.605% including 1.626 for the total power forecast
CNN-M-BDLSTM [27]	CNN with multilayer bidirectional LSTM	Residential power consumption data such as active power, voltage, current and time	Short-term rodential power consumption prediction	MSE: 0.3489 RMSE: 0.5905 MAE: 0.3730
AE-LSTM [28]	Auto encoder with long short-term memory	Daily power generation, max grid connected power, radiance	Short-term solar power generation prediction	AE-LSTM with an MAE of 0.05–0.09. MAPE was not calculated
CNN+A-LSTM+AR [29]	Integrates CNN, attention-based LSTM, and autoregressive	Power generation data from different renewable energy sources and weather data	Forecasting power generation of multiple renewable energy sources (solar PV, solar thermal, wind)	MAPE: 9.16% (solar PV), 18.15% (solar thermal), 16.87% (wind)
CNN-Bi-LSTM [30]	Combines CNN and bidirectional LSTM	Renewable energy data (solar PV, wind) and weather data	Forecasting renewable energy power generation (Solar PV, Wind)	CNN-Bi-LSTM has an error of 00219 with solar data and 1.0125 with varying wind data
Bi-LSTM [32]	Bidirectional long short-term memory network	Hourly power generation data from 20 MW PV plant	Predicting solar power generation from large-scale photovoltaic (PV) plants	R value: 0.98 (one-hour ahead predictions)
CNN-Bi-GRU [33]	Multimodal information fusion (time series + textual data)	Historical renewable power generation data, policy documents	Short- and long-term renewable electricity demand forecasting	Lower RMSE and MAPE compared to ARIMA, standalone GRU, EEMD-ARIMA
CNN-LSTM [31]	Sequence-to-sequence hybrid model (CNN as encoder, LSTM as decoder)	Monthly peak load data (New South Wales, Australia)	Long-term monthly peak load forecasting (three-year horizon)	MAPE: 4.29% (36-month ahead)
Proposed Framework	Combine Multiple Model (Best model based on Source) in one Framework	Historical monthly EIA data (USA’s electricity production by source)	Forecasting USA’s Power Generation for Fossils Fuel and Renewable Sources	MAPE for Nuclear: 1.66% Details in Table 4

. In this section, the methodologies, performance, and the applicability of various contemporary approaches have been discussed, including STL-CNN [26], CNN-M-BDLSTM [27], AE LSTM [28], CNN-A-LSTM-AR [29], CNN-Bi-LSTM [30], Bi-LSTM [32], CNN Bi GRU [33], CNN-LSTM [31].

Compared to prior studies that focused on single-model architectures or specific energy types, the proposed framework introduces a novel multi-model approach by selecting the best-performing model for each generation source and integrating them into a unified forecasting system. While earlier works reported strong performance in specialized domains—such as STL-LSTM achieving a MAPE as low as 1.626% for total power in Turkiye [26] or AE-LSTM obtaining MAE between 0.05 and 0.09 for solar prediction [28]— they often targeted limited sources or regions. Other hybrid models like CNN-Bi-LSTM [30], CNN-LSTM [31], and CNN-Bi-GRU [33] demonstrated competitive error rates but were constrained to renewable datasets or short-term horizons. In contrast, our framework delivers highly accurate forecasts across both fossil and renewable sources using comprehensive historical EIA data, achieving, for example, a MAPE of 1.66% for nuclear generation, with detailed performance metrics presented in Table 8. This broad, source-specific optimization makes it better suited for national-scale energy planning and policy applications.

5. Conclusions

Our findings demonstrate that the proposed multi-model deep learning framework, selecting the best-performing architecture for each individual energy source, can play a pivotal role in strategic power generation planning, particularly as the grid transitions toward decarbonization and higher renewable energy integration. This source-specific optimization significantly improves forecasting accuracy compared to single-model or traditional approaches, as evidenced by superior MAPE, RMSE, and MAE results across both fossil fuel and renewable categories. By accurately forecasting power generation from diverse sources, the framework provides critical insights for balancing supply and demand, optimizing resource allocation, and supporting policy decisions. The ability to adapt to non-linear patterns in complex energy datasets makes it especially valuable in managing the increasing variability introduced by renewables such as solar and wind. Furthermore, the integration of STL decomposition further enhances the framework’s ability to capture long-term trends, seasonal behaviors, and short-term variability across diverse generation sources, an area where many existing models fall short.

As renewable penetration grows, such data-driven, source-tailored forecasting frameworks will be essential for ensuring grid stability, enhancing energy efficiency, and achieving long-term sustainability goals. For future work, we plan to extend the framework into a multimodal large language model (LLM)-based system capable of integrating textual reports, policy documents, and historical datasets. This enhancement will allow the framework to capture the dynamic influence of policy shifts and regulatory changes on power generation, making it even more adaptable to real-world operational and planning challenges.

Although electricity generation is influenced by exogenous variables such as meteorological conditions such as fuel prices, economic activity, and policy interventions, this study adopts a univariate forecasting framework to evaluate the intrinsic capability of the proposed hybrid deep learning models to capture temporal patterns directly from historical generation data as consistent long-term monthly records for all relevant exogenous variables across multiple energy sources are not uniformly available. Incorporating variables such as temperature, wind speed, solar irradiance could potentially improve forecasting performance, especially for renewable sources. Future work will extend this framework to a multivariate setting.