Enhancing Day-ahead PV Power Forecast Accuracy with an Improved Informer Model Using AdaBelief Optimization

Renewable Energy sources (REs) have recently emerged as crucial elements in the sustainable development of human society. Therefore, an accurate REs power generation forecast is considered to be a key countermeasure to maintain the stability and reliability of the grid. Recently, advanced Transformer-based models have gained more and more attention from researchers, especially in the field of time-series forecasting. This paper proposed an Informer model improved with AdaBelief optimizer for day-ahead PV power forecast. Replacing the conventional self-attention with ProbSpare self-attention technique increases the training efficiency and forecasting accuracy and significantly reduces the training period compared to the basic Transformer. Moreover, the AdaBelief, can both bear faster convergence and effective generalization, further enhancing the precision of the proposed model. A 5-minute multivariate dataset is employed for accessing the forecasting performance of the proposed model and demonstrates the enhancement of each module to the forecasting efficiency.