Adaptive LSR fusion of multi-scale 3-branch CNN-BiLSTM, LSTM and CNN-BiLSTM models for robust multi-seasonal and multi-horizon weather forecasting in renewable energy management

Abstract

Accurate and robust forecast of meteorological variables such as wind speed, solar irradiance and ambient temperature is challenging because of their nonlinear, non-stationary behavior as well as season-dependent dynamics. This study proposes an adaptive least squares regression fusion-based ensemble method for multi-horizon multi-season weather forecasting by combining three individual deep learning models: a proposed multi-scale three-branch convolutional neural network with bidirectional long short-term memory, a classical long short-term memory, and convolutional neural network with bidirectional long short-term memory. The least squares regression fusion adaptively assigns the appropriate weights to each individual model according to the weather variable, season, and forecast horizon. Evaluation tests conducted on a one-year dataset for different short-term horizons reveal the superior performance of least squares regression fusion compared to all individual deep learning models in terms of accuracy and performance stability. For instance, at one-hour forecasting horizon, the outcomes show, a root mean square error reduction ranged from 3–12% for irradiance, 3–10% for wind speed and 4–10% for temperature. Additionally, a high coefficient of determination was observed, approximately equal to 0.99, implying a strong temporal correlation between the predicted and observed weather variables throughout the four seasons. Statistical analyses, including paired t-tests with false discovery rate correction, confirm that least squares regression fusion consistently outperforms individual models, achieving the highest win rates for wind speed (65.8%) and irradiance (55.8%), while remaining competitive for temperature (46.4%). Overall, the adaptive least squares regression fusion framework effectively integrates heterogeneous deep learning models, dynamically adjusting their corresponding contribution, and achieves an effective forecast of weather variables for short-term multi-horizons and across all four seasons. • Multi-season, multi-horizon short-term forecasting of three weather variables. • Adaptive LSR-based fusion of MS-3B-CNN-BiLSTM, LSTM, and CNN-BiLSTM with chronological train–test splits. • LSR fusion outperforms individual DL models across all seasons and forecast horizons. • Statistical validation using paired t-tests with false discovery rate correction confirms the superiority of LSR fusion. • Robustness and high accuracy of the model are achieved specially for temperature and irradiation.