Improved multi-gate mixture-of-experts framework for multi-step prediction of gas load

Abstract

Accurate prediction of gas load is critically important for the stable gas usage and accurate dispatch. In the existing literatures, the prediction accuracy is limited due to the fact that the statistical methods only consider the linear relationships between the gas load and the data-driven seq2seq neural network, which suffers from model compression with shape-based loss function. To address the above issues, an improved multi-gate mixture-of-experts framework is put forward. Firstly, the relevant non-temporal features are selected by analyzing the change pattern of gas consumption, and the Boruta algorithm is used to screen the irrelevance and redundancy features. Secondly, convolution network, gated recurrent unit networks and auto regression are chosen as expert networks to acquire both short-term and long-term temporal features, which will be fed into gated network and tower network to achieve multi-step prediction. Finally, the Dilate loss function is used to learn the optimal weight of the designed model considering the dynamics of both shape and temporal. Multi-step prediction experiments with the real constructed gas load dataset verify the effectiveness of the proposed approach, and the evaluation metrics from two perspectives of traditional regression and gas load dispatching verify that the improved multi-gate mixture-of-experts outperforms the state-of-the-art methods.