Long-Short Step Coordinated Photovoltaic Power Generation Fusion Forecasting

Accurate photovoltaic (PV) power generation forecasting is crucial for the stability and scheduling efficiency of the smart grid. However, the high volatility and non-stationarity of solar irradiance lead to continuous error accumulation when predicting over longer time horizons, posing a significant challenge for long-term forecasting. To address this issue, this paper proposes a novel hybrid framework integrating iTransformer, Mixture-of-Experts (MoE), and Bidirectional Long Short-Term Memory (BiLSTM) network. Specifically, the iTransformer is employed to capture global temporal dependencies. Internally, a gating mechanism replaces the traditional feedforward network with a MoE layer, enabling dynamic adaptation to diverse weather patterns (e.g., sunny vs. cloudy days). Subsequently, a BiLSTM layer is utilized to optimize local temporal features and smooth the predicted sequence. Furthermore, to enhance the stability of longterm forecasting, we introduce a Long-Short Timestep Collaborative Forecasting Strategy. This strategy corrects the training loss by calculating the high-precision residuals of smallscale temporal predictions, thereby guiding the optimization process of long-term forecasts and mitigating performance degradation caused by long sequences. Experimental results demonstrate that the proposed method effectively reduces forecasting errors under different weather conditions and exhibits superior forecasting performance in terms of both accuracy and stability.