超短期LSTM风电功率预测模型的混合专家模块化代理解释方法

Modular Surrogate Interpretation Method Based on Decision Tree Mixture of Experts for Ultra-Short-Term LSTM Wind Power Forecasting Model

深度学习模型已被广泛应用于超短期风电功率预测。对黑盒深度学习模型预测过程的决策逻辑进行解释和分析,有利于提升预测模型和预测结果的可信度,减小不确定性造成的电力系统运行风险。对此,文章针对经典的长短期记忆网络(long short-term memory,LSTM)超短期风电功率预测模型,提出了一种基于决策树混合专家模型(decision tree mixture of experts,DTMOE)的模块化代理解释方法。将LSTM模型内部的预测过程分解为两个相对独立的模块,采用DTMOE分别对两个模块的输入输出进行拟合,通过分析DTMOE的拟合结果对LSTM模型的预测过程和逻辑进行映射解析。算例分析表明,DTMOE模型对原始黑盒模型有较高的拟合精度与可解释性能力;DTMOE模型的可视化结果可以解析和展现LSTM模型预测过程的决策路径以及关键影响特征。

Deep learning models have been widely used for ultra-short-term wind power forecasting.The decision logic of the forecasting process in deep learning models can be explained and analyzed to improve the credibility of the forecasting models and forecasting results,and to reduce the operational risks of power systems caused by uncertainties.A modular surrogate interpretation method based on the decision tree mixture of experts(DTMOE)is proposed for the classic long short-term memory(LSTM)model used in ultra-short-term wind power forecasting.The forecasting process of LSTM is divided into two relatively independent modules,each of which is fitted separately using DTMOE to fit its input and output.Therefore,the forecasting process and logic of LSTM are mapped and interpreted by analyzing the fitting results of DTMOE.Case studies show that the DTMOE model has higher fitting accuracy and interpretability than the LSTM model.The decision paths and key influential features of the forecasting process created by LSTM can be revealed and displayed through the visual results of the DTMOE.