基于PSR-LSTM的机组负荷短期预测研究

Study on Short-term Load Forecasting of Units based on PSR-LSTM

为了提高机组负荷短期预测精度,针对其非线性、时序性特点,以某660 MW机组为研究对象,提出一种基于相空间重构(PSR)和长短期记忆网络(LSTM)的负荷预测模型PSR-LSTM。利用归一化函数(mapminmax)将原始机组负荷数据归一化处理后,选用C-C法与小数据量法证明历史负荷数据具有混沌特性并进行负荷时序重构;将重构后的每一维特征向量作为时间步输入建立的LSTM模型训练进行短期预测。研究表明:PSR-LSTM预测模型在12 h与在5 min内的平均绝对百分比误差分别为1.38%和0.39%,均方根误差分别为6.38和1.83;相较于标准LSTM模型以及传统自回归滑动平均模型(ARMA),PSR-LSTM模型误差较低并具有更高的预测精度。

In order to improve the accuracy of short-term load forecasting, a PSR-LSTM load forecasting model based on phase space reconstruction(PSR) and long-term memory network(LSTM) is proposed for a 660 MW unit.After normalizing the original unit load data with the normalized function(mapminmax),C-C method and small data method are used to prove the chaotic characteristics of historical load data, and load time series reconstruction is carried out;each dimension feature vector after reconstruction is used as time step input to establish LSTM model for short-term prediction.The results show that: the average absolute percentage error of PSR-LSTM prediction model in 12 h and 5 min is 1.38% and 0.39%,and the root mean square error is 6.38 and 1.83,respectively.Compared with the standard LSTM model and the traditional auto-regressive moving average model(ARMA),the model error is lower and the prediction accuracy is higher.