基于LSTM的高校建筑电力负荷预测方法

满   达 张卓凡 张金金 谢将剑（北京林业大学，北京市 100083）

Power Load Forecast Method for University Buildings Based on LSTM

MAN Da ZHANG Zhuofan ZHANG Jinjin XIE Jiangjian

(Beijing Forestry University,Beijing 100083，China)

Abstract：A forecast model based on long short term memory(LSTM)is proposed to improve the accuracy of power load forecast in university. Taking the power load data of a university as the research object,this paper analyzes the influence of different hyper‑parameters,determines the optimal forecast model,and compares it with the commonly used load forecast model based on support vector machine(SVM). The results show that the mean absolute percentage error(MAPE)of the load forecast model proposed in this paper is 6.67 % for office buildings,4.32 % for scientific research buildings,5.98 % for teaching buildings and 4.57 % for dormitories. In addition, each type of building is about 1.5 % lower than the MAPE based on SVM forecast model.

Key words：university buildings;load forecast；

load characteristics；load curve;LSTM;support vector machine;mean absolute percentage error;experimental comparison

摘　要：提出一种基于长短时记忆网络（LSTM）的预测模型，以提高高校电力负荷预测的准确率。以某高校的电力负荷数据为研究对象，分析不同超参数的影响，确定最优的预测模型，并与常用的基于支持向量机（SVM）的负荷预测模型进行负荷预测对比。结果表明，本文提出的负荷预测模型平均绝对百分误差(MAPE)为：办公楼6.67 %、科研楼4.32 %、教学楼5.98 %和宿舍4.57 %，每类建筑均比基于SVM预测模型的MAPE低1.5 %左右。

关键词：高校建筑；负荷预测；负荷特征；负荷曲线；长短时记忆网络；支持向量机；平均绝对百分误差；实验对比

中图分类号：TU852      　　  　  文献标识码：A

doi：10.3969 / j. issn.1003 － 8493.2021.11.013