Multi-Task Learning Based Joint Forecasting of Wind Power Generation and Power Load

Abstract

Wind power generation and electricity load are influenced by common meteorological and temporal factors, yet conventional methods generally forecast them independently. This paper proposes a multi-task learning framework combining a shared encoder with two task-specific attention branches for joint wind-power and load forecasting. Experiments are conducted on a renewable-energy and electricity-demand dataset released by the California Independent System Operator and the National Renewable Energy Laboratory, with the original 5-minute observations resampled to 15-minute intervals. Direct normal irradiance, diffuse horizontal irradiance, global horizontal irradiance, wind speed, humidity, temperature, temporal variables, and historical target observations are used as inputs. The framework is compared with eight conventional, single-task, and joint deep-learning baselines using mean absolute error, root mean squared error, mean absolute percentage error, and the coefficient of determination. It achieves mean absolute error/root mean squared error values of $54.017 / 96.842 \mathrm{MW}$ for windpower forecasting and $86.925 / 120.318 \mathrm{MW}$ for load forecasting, outperforming the strongest baseline on both tasks. The results demonstrate the benefit of joint representation learning, while component-wise ablation confirms the independent contributions of shared encoding and task-specific attention.