基于长短期记忆网络的短期空调冷负荷预测
Short-term air conditioning cooling load prediction based on long-short-term memory network
摘要：
提出了一种基于长短期记忆网络（LSTM）的短期空调冷负荷预测模型，仅采用历史负荷数据预测未来1 d的逐时冷负荷。通过与传统的BP神经网络模型进行对比，验证其准确性。为了进一步提高模型预测精度，对网络结构（包括输入层、输出层及隐含层神经元数量）与预测策略进行了优化，获得最优的预测模型。结果表明，基于LSTM的预测模型可实现准确的负荷预测，且与BP神经网络模型相比，预测精度更高，均方根误差和均方根误差的变异系数分别降低116 kW和5.42%。对LSTM模型优化的结果表明：利用历史7 d负荷数据预测未来1 d的逐时空调负荷是最佳的输入输出组合选择；隐含层神经元数量为60时，模型精度较高且较为稳定；采用分步输出的预测策略能降低峰值负荷时的预测误差，提高负荷预测精度。
Abstract：
This paper proposes a short-term air conditioning cooling load prediction model based on the long-short-term memory network (LSTM), which only uses historical load data to predict hourly cooling load in the next day. By comparing with the back-propagation neural network (BPNN) model, the accuracy of the model is testified. In order to further improve the prediction accuracy of the model, this study optimizes the network structure including the number of neurons in input layer, output layer and hidden layer and prediction strategies to obtain the optimal prediction model. The results show that the prediction model based on the LSTM can forecast cooling load accurately, and perform better than BPNN model, and the root mean squared error and its coefficient of variation reduce 116 kW and 5.42%, respectively. The optimized results show that the best combination of input and output is to use historical seven-day load data to predict the hourly air conditioning load of the next day. When the number of neurons in the hidden layer is 60, the prediction accuracy of the model is higher and more stable. The stepwise output prediction strategy can reduce the prediction error at the peak load and is helpful to improve the prediction accuracy.
Keywords:airconditioning;coolingloadprediction;predictionmodel;long-short-termmemorynetwork;neuralnetwork;neuron;rootmeansquarederror