Research On Power Load Prediction Based On Time Series Convolution Lstm

In modern power system,power load forecasting has been an indispensable and important work in the development of power industry.Accurate power load forecasting is the basis of safe,stable and economical operation of power system,reasonable load forecasting provides an important basis for the decision-making of power grid dispatching,operation and resonable pricing of power system.In the long term,the long-term planning and construction of the whole power system are closely related to power load forecasting.In power load forecasting,the significance of super short-term load forecasting mainly lies in online control and real-time monitoring of power system,prevention of potential accident risk and formulation of reasonable power generation plan.Accurate prediction of power load is not only important for the efficient,stable and safe operation of power system,but also has a great impact on the production and life of the whole society.Therefore,it is an important work to improve the accuracy of power load forecasting.With the development of deep learning,neural network has been widely applied in the field of time series prediction,including a large number of researches related to power load prediction.In many neural network models,RNN is widely used because of its memory ability,the standard RNN model can't learn long time dependence because of the gradient disappearing,the emergence of LSTM solves this problem.The "gate mechanism" of LSTM solves the problem that RNN model cannot learn long-term dependence,and is widely used in the field of time series prediction.This paper is based on LSTM model and improves LSTM model,and applies it to the field of super short term power load forecasting.The main work of this paper is as follows:1.A time-series convolution LSTM model based on one-dimensional convolution is proposed.Aiming at the problem of insufficient ability of LSTM to extract high-dimensional features and multi-scale features of time series data,improve the link of LSTM feature extraction by using convolution operation.By introducing multi-layer convolution of causal convolution and dilated Convolution,higher dimensional and multi-scale features in time series are extracted.In the process of multi-layer convolution operation,combined with the ability of LSTM to capture time series dependence in time series data,LSTM models were established to train the hierarchical dimension features extracted from the convolution operation.By dynamically integrating the convolution operation with LSTM,the model can learn more dimensions and scales of time series features and LSTM can obtain longer data dependence,so as to enhance the prediction accuracy of the model.2.A multi-channel structure model is designed.LSTM models usually only have a single input sequence and cannot capture the periodic features in power load data.Therefore,a multi-channel structure model is designed to train the model by receiving multiple input sequences at multiple time scales.3.Construct multi-periodic scale input sequence.Aiming at the periodic characteristics of power load data,the adjacent time feature sequence and periodic time feature sequence are extracted from the original sequence respectively,and the input sequence of the model is constructed from three dimensions of adjacent time,short period and long period for training.4.The validity of the above models and methods is verified by experiments.The time-series convolution LSTM model is established,and experiments are carried out with time-series convolution LSTM model on the power load data of two different regions,and the experimental results are compared with other models.The results show that the time-series convolution LSTM model has higher prediction accuracy.The time-series convolution LSTM model with multi-channel structure is established,and the input sequence of multi-period scale is constructed as the input of the model.Experimental results show that the prediction accuracy of the model is higher than that of the single time-series convolution LSTM model.