Spatial-Temporal Graph Hybrid Neural Network for PV Power Forecast

Amid the bloom of Renewable energy sources (RES) integration into the electrical grid, an accurate power generation forecast is essential for a proper energy management plan to ensure power system reliability. Nevertheless, most previous studies have concentrated on predictors designed for centralized PV farms, often falling short of effectively accommodating distributed PV systems. Hence, this paper proposes a Graph- based architecture, namely Graph Convolutional Long Shortterm Memory Neural Network (GCLNN) for multi-site PV power forecast. Using the Graph Attention Mechanism, the adjacency matrix is constructed, representing the correlation between variables. From that, the weight of each parameter of the input data is determined, takes part in the training process, and improves the forecasting accuracy. For a practical indication of the proposed model's superiority, a multi-site dataset consists of the historical output power of three different neighboring PV systems in Ha Tinh, Vietnam. From the experimental result, the proposed model brings out outstanding accuracy, especially in higher time-step forecasts compared to the benchmark models.