Multi‐Step Ahead Short‐Term Load Forecasting Based on MCNN‐MMoL

ABSTRACT Multi‐step ahead forecasting of power load is crucial for optimising power system scheduling and participating in energy market transactions. However, existing forecasting methods often suffer from issues such as cumulative prediction errors and insufficient modelling of sequence dependencies. To solve the above problems, a multi‐step ahead forecasting method based on multiplexed convolutional neural networks (MCNN) and multi‐gate mixture of long short‐term memory networks (MMoL). First, multi‐branch convolution is adopted to construct independent feature spaces for different levels of power loads, achieving multi‐scale feature fusion to enhance the representation ability of the input samples. Next, the multi‐step prediction task is transformed into a multi‐task joint optimisation problem. Multiple independent LSTMs are used as shared experts, and task‐specific gating units are utilised to dynamically learn the optimal combination of expert models for each future time step, achieving more refined time‐series feature modelling. Finally, comparative experiments are conducted based on two real‐world datasets. The results show that the proposed model exhibits better accuracy and robustness.