Short-term CHP heat load forecast method based on concatenated LSTMs

A concatenated LSTM (Long Short-Term Memory) architecture for CHP (combined heat and power) heat load forecasting was presented. Firstly, input data was normalized and separated into historical climate and heat load data. Then feed the separated data into two LSTM neural networks. Finally, the two LSTM models were concatenated as inputs to another LSTM model followed by two dense layers. Relu function is used as activation function for the dense layers and ADAM (Adaptive moment) method was used as the gradient based optimizer. The concatenated LSTM architecture was trained and tested on heat load data from Nov.2016 to Feb.2017 of Rizhao, Shandong. Experimental results show an obvious improvement in the forecasting accuracy compared with simple LSTM.