Accurate short-term photovoltaic power forecasting remains a significant challenge for grid operators managing high penetrations of solar generation. Weather variability, cloud cover, and seasonal changes create rapid fluctuations in PV output that traditional forecasting methods struggle to capture. Researchers have now developed an advanced hybrid framework that combines multiple artificial intelligence techniques to address these limitations.
The proposed system uses three primary components working in concert. First, ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) breaks down historical solar generation data into component signals, reducing the complexity of nonlinear patterns. These components are then organized into entropy-based groups to filter out noise and emphasize meaningful signals. Second, a bidirectional long short-term memory (BiLSTM) neural network with self-attention mechanisms learns temporal patterns across extended time periods, capturing both recent trends and longer-term dependencies in the data. Finally, an improved mantis shrimp optimization algorithm fine-tunes model parameters automatically for each data component.
Testing against real operational data from China's State Grid Corporation demonstrated substantial improvements. The combined model reduced root mean square error (RMSE) by 9.37% and mean absolute error (MAE) by 13.31% compared to baseline approaches using fixed parameters. Ablation studies confirmed that each component—particularly the entropy-based regrouping step—contributed meaningfully to performance gains.
These improvements have direct implications for grid reliability and economic efficiency. Better PV forecasts enable grid operators to schedule conventional generation more accurately, reduce reserve margins, and minimize costly ramping events. As solar deployment accelerates globally, forecasting accuracy becomes increasingly important for maintaining grid stability and maximizing renewable energy penetration without compromising system reliability.
