Optimized design and implementation of a power grid load forecasting model for renewable energy integration scenarios

Abstract

Given the challenges associated with renewables-based power grids, such as greater volatility, greater non-stationarity, and traditional forecasting techniques failing to sufficiently adjust to adaptive changes, this paper studies the design and implementation of an adaptive power grid load forecasting model suitable to integrate renewables in power grids. We used multi-source load data as the basis of data engineering and pre-processed the data to create a unified load, photovoltaic (PV), wind, meteorology and time encoded data vector. We designed a load forecasting model to use a combination of dilated convolution, Transformer encoding, and encapsulated perturbation gating, as well as an optimized training approach based on a weighted loss function, adaptive dynamic learning rates, and residual correction. Our approach achieved best results when compared to other models in mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), as well as provided stable forecasting results in high PV penetration forecasting conditions, forecasting conditions with sudden changes in wind, and meteorological conditions forecasting with high volatility. This study provided a multi-source time-series forecasting method.