Attention Mechanism Based Probabilistic Day-Ahead Net-Load Forecasting with Behind-the-Meter Solar

The increasing penetration of behind-the-meter solar photovoltaic is deteriorating the accuracy of legacy load forecasting algorithms, disrupting the operation and planning of power systems. Moreover, the specifications of the behind-the-meter photovoltaic systems are not well exposed to the grid operators, which is further complicated by the possible presence of energy storage systems. In this paper, an attention mechanism based graph transformer network, which utilizes the information from the neighboring zones and accounts for the temporal correlations within a time series, is proposed to perform probabilistic day-ahead net-load forecasting. The proposed architecture forecasts the net load directly without the need of knowing the capacity of solar-plus-storage or historical solar generation. Numerical results on the data from Open Power System Data show superior performance over six baselines.