The Lightweight Hybrid Deep Learning Approach for Capturing Long-Term and Short-Term Constraints for an Accurate Solar Radiation Forecast

A Hybrid Deep Learning–XGBoost Framework for Short-Term Solar Power Forecasting to Enable Sustainable Smart Grid Integration

MSF-TransPV: a multi-source fusion transformer for short-term photovoltaic power forecasting

Accurate short-term photovoltaic (PV) power forecasting is essential for secure grid operation, economic dispatch, and efficient utilization of renewable energy. However, PV output exhibits strong nonlinearity and nonstationarity due to rapidly varying meteorological conditions, cloud movements, and system uncertainties. Conventional statistical models and single-source deep learning approaches often fail to fully exploit the rich multi-source information available in modern PV plants, such as historical power, on-site meteorological measurements, and clear-sky or numerical weather prediction (NWP) features. In this paper, we propose MSF-TransPV, a Multi-Source Fusion Transformer framework for short-term PV power forecasting. The model adopts a multi-branch temporal encoder that separately processes historical PV output, meteorological variables, and optional clear-sky/NWP-derived features, mapping them into a shared latent space. A cross-variable multi-head attention module is then introduced to explicitly capture the dependencies between PV dynamics and atmospheric conditions, enabling fine-grained interaction across different feature sources. On top of the fused representation, a temporal Transformer encoder models long-range temporal dependencies and feeds a lightweight decoder that supports both point forecasting and probabilistic forecasting via quantile regression. Experimental results on real-world PV datasets demonstrate that MSF-TransPV consistently outperforms persistence, statistical baselines, recurrent neural networks, and vanilla Transformer models in terms of RMSE, MAE, and normalized error metrics, while also providing reliable prediction intervals. These results indicate that explicit multi-source fusion and cross-variable attention are highly effective for PV forecasting under complex and rapidly changing weather conditions.

Solar Energy Production Forecasting With Hyperparameter-Optimized Time-Series Dense Encoder

Photovoltaic power production forecasting is essential for grid stability, market operation, and large-scale integration of renewable energy, yet it remains challenging due to the nonlinear and weather-driven nature of solar generation. This work introduces a fully automatic photovoltaic forecasting framework that combines the time-series dense encoder architecture with a genetic hyperparameter space explorer based on the non-dominated sorting genetic algorithm III. The proposed framework employs a computationally efficient encoder-decoder architecture that leverages historical photovoltaic production, sun-position features, and meteorological covariates, while evolutionary search adapts architectural and training configurations to each forecasting horizon. The approach is evaluated on a six-year real-world dataset from a solar farm in Colombia, covering short-, medium-, and long-term horizons. Across all experiments, the proposed method achieves strong results compared to statistical, machine learning, and deep learning baselines. In addition, a rigorous statistical evaluation provides detailed insight into the reliability and robustness of the forecasts. A complementary computational analysis quantifies training and inference complexity, showing that the proposed approach is computationally efficient and scalable. These results show that the proposed framework provides a promising approach for solar power forecasting, with potential to support energy planning, resource allocation, and the broader integration of renewable energy technologies.

Multi-output deep learning framework for joint forecasting of solar irradiance and wind speed with cross-regional transferability analysis

Accurate forecasting of solar irradiance and wind speed is essential for improving hybrid renewable energy systems and ensuring grid stability. This study develops and evaluates a multi-output deep learning framework for the simultaneous prediction of global horizontal irradiance (GHI) and wind speed across multiple Indian regions. Hourly data from the National Solar Radiation Database (NSRDB) for the period 2015–2020 were used to train light gradient boosting machine (LightGBM), long short-term memory (LSTM), bidirectional long short-term memory (BiLSTM), and convolutional long short-term memory (ConvLSTM) models, with cross-regional transfer learning applied across Tamil Nadu, Kerala, Karnataka, and Andhra Pradesh. Among the models, ConvLSTM achieved the best performance with a mean absolute error (MAE) of 0.061 and an R² value of approximately 0.91, while BiLSTM demonstrated comparable accuracy with lower computational cost. The proposed framework emphasizes cross-regional transferability, demonstrating robust generalization across heterogeneous climatic conditions. Error distribution analysis further indicates improved prediction stability, with ConvLSTM exhibiting lower variability compared to other models. These results support scalable and reliable renewable energy forecasting, with practical implications for grid operation, power electronic control, and hybrid energy system management.

A Data‐Driven Investigation of How Forecast Horizon and Training Data Size Influence Machine Learning Performance in Solar Irradiance Forecasting

Accurate short‐ and medium‐term solar irradiance forecasting is vital for integrating solar power into the grid, but high variability and weather dependence make it challenging. Machine learning (ML) models offer promise, but their performance often depends on forecast horizon and training data size. In this study, four models, including random forest (RF), support vector regression (SVR), gradient boosting regressor (GBR), and deep neural networks (DNN), are compared for 15‐min, hourly, and day‐ahead solar irradiance prediction. The models are trained with meteorological and irradiance data for the city of Sioux Falls, South Dakota, using datasets ranging from 1 to 8 years (2012–2019). A total of 96 models are developed by combining four algorithms across three horizons and eight training data sizes. Using the maximum relevance minimum redundancy (MRMR) technique, the best features are selected, and the results are evaluated using standard error metrics. The results indicate that RF provides the most stable accuracy prediction, SVR provides the best accuracy on day‐ahead forecasts with limited data, GBR performs best at the hourly prediction level, and DNN accuracy is more sensitive to both data quality and overfitting. It emphasizes the importance of choosing models that align with the forecast horizon and the size of available data. In this paper, we provide an in‐depth comparison of four state‐of‐the‐art ML models across various horizons and training data sizes. Results show the importance of integrating MRMR, with guidelines for researchers and energy planners to achieve more accurate and efficient solar irradiance prediction.

A TCN-Attention Model for High-Accuracy Solar Energy Generation Forecasting

Predictions of solar power output that are close to the mark may help with grid stability, energy management, and the integration of renewable energy sources. Prediction using conventional statistical models is challenging due to the substantial influence of nonlinear meteorological variables and temporal variations on solar power output. A deep learning-based forecasting framework integrating an attention mechanism and a Temporal Convolutional Network (TCN) is proposed in this study to enhance the precision of solar energy prediction. A number of feature engineering and pre-processing approaches are used by the model, including lag features, rolling statistics, temporal indicators, and outlier reduction using the Interquartile Range (IQR) methodology. To store long-range temporal associations, the TCN architecture employs dilated causal convolutions; the attention mechanism highlights the most important time steps in the input sequence. The proposed model outperforms state-of-the-art baseline models such as Long Short-Term Memory (LSTM) and traditional TCN when tested on a publicly available solar energy dataset. It can accomplish 253.32 MW Mean Absolute Error (MAE), 325.04 MW Root Mean Square Error (RMSE), 2.30% Mean Absolute Percentage Error (MAPE), and 0.9804 Coefficient of Determination (R2). These parameters are evaluated by experimental means.