A Channel-Wise Cross Attention Mechanism Based Aggregated Distributed Photovoltaic Power Day-ahead Forecasting Method

To enhance the safety, stability, and economic efficiency of distribution networks, improving the day-ahead forecasting accuracy of distributed photovoltaic (PV) power plants is of significant importance. Given that PV output is influenced by complex factors such as weather, climate, and seasons, traditional forecasting methods based on irradiance forecasts and single-channel historical data struggle to handle the diverse operational scenarios of PV plants. This paper proposes a forecasting method for distributed PV clusters based on a channel-wise cross-attention mechanism. First, the cross-attention mechanism adaptively learns the pattern correlations among different channels of distributed PV data. Building on this, a forecasting model with a time-series Transformer backbone is employed to achieve spatially complementary pattern matching for distributed PV predictions and to reduce overall model prediction errors. Finally, the effectiveness of the proposed method is validated through test cases, demonstrating the substantial potential of spatio-temporal correlations in distributed PV for improving forecasting accuracy.