| With the gradual expansion of the scale of heating,the central heating system in the heating process often occurs that the temperature adjustment time is too long,the supply exceeds the demand or the demand exceeds the supply,resulting in the increase of system operating costs.Therefore,the realization of need-based heating and the provision of technical support for the stable operation of the central heating system has become an important problem to be solved urgently in the industry.With the development of load prediction technology and the update of deep learning methods,more and more prediction methods have been proposed.Reliable heat load prediction has also become the focus of research in the field of heating.How to effectively extract the deep characteristics of heat load data is an important problem of heat load prediction.This paper takes a deep research on the heat load prediction,which is based on a central heating system in Hebei province.Following the progressive research idea,this paper studies the feature extraction of heat load data and the optimization of prediction algorithm to provide reliable theoretical guidance for the stable operation of central heating system.The main work of this paper is as follows:(1)Time feature fusion LSTM model is proposed.The model is improved from two aspects.Firstly,the heat load data was analyzed to extract the proximity,periodicity and trend characteristics of the data.Secondly,with the idea of integrated learning,LSTM network is used to predict the proximity,periodicity and trend data respectively,and then the three prediction results are combined with external characteristics through LSTM network to obtain the final prediction results.Where,the external feature used in fusion is the outdoor temperature at the prediction time,and the LSTM network which build the prediction model consists of a layer of LSTM and a layer of RNN.Through the improvement of the above two points,the prediction accuracy of model is improved.In the case of insufficient experimental data,the model provides an effective solution.(2)A hybrid spatial-temporal genetic algorithm optimization for attention-LSTM is proposed.This model is based on the time feature fusion LSTM model,and it is improved from three aspects.Firstly,in order to fully consider the influence factors and improve the prediction accuracy of the model,the data from adjacent heat exchange stations are introduced as the spatial characteristics of target heat exchange stations.Secondly,as the input sequence length increases,errors will gradually accumulate in the LSTM network,which leads to LSTM inability to properly process longer-term time series data.Therefore,attention mechanism is used to improve the LSTM network,improve the sensitivity of the model to information,and realize the selective attention to features.Thirdly,genetic algorithm is used to select the parameters of the attention-LSTM network to reduce the impact of random selection of parameters on the model.Where,the structure of LSTM network in this model is the same as that in time feature fusion LSTM model.Through the improvement of the above three points,the prediction accuracy of the model is further improved.When the experimental data are sufficient,the model provides a more perfect prediction method. |
