Research On Power Load Forecasting Based On Spark And Deep Learning Model

Electric energy is one of the indispensable energy sources in our life.How to improve the efficiency of electricity consumption and make the plan of the quantity of electricity rationally have become one of the focuses of people’s attention.Accurate power load forecasting is not only an important and effective energy-saving measure,but also has a huge positive impact on economic development,policy formulation,and social security.This paper aims to conduct power load forecasting research based on Spark and deep learning techniques.The specific research work is as follow:(1)The long-term electric load forecasting model based on "year",the medium-term electric load forecasting model in "month" and the short-term electric load forecasting model in "hours" are proposed respectively.For the features of the three types of power load data,the three-layer LSTM model predicts monthly electricity consumption,the LSTM and MLP combined models predict annual electricity consumption,and the multivariate deep encoder-decoder model predicts hourly power usage have been proposed separately.It is proved by experiments that the three models proposed in this paper have better performance in the respective data sets than the benchmark models.(2)The LSTM model optimization method is proposed.By decomposing the components of the LSTM gradient into components produced by different paths,mathematical reasoning is used to prove that the linear time path(cell state)contains information about long-term dependencies,and when the LSTM weight is relatively large,the gradient through the path is suppressed.It is proved by experiments that the LSTM optimization model proposed in this paper has better performance in short-term electric load forecasting in "hours" than the LSTM model and short-term electric load forecasting model.(3)A parallelization implementation method for short-term electric load forecasting model is proposed.According to the features of short-term electric load data and prediction model,a data compression storage scheme based on HDFS and Spark RDD and an asynchronous stochastic gradient sharing scheme are proposed,solving the parallel problem of short-term electric load forecasting model.After in-depth study of the Spark and DL4 J frameworks,achieving parallelization of short-term power load forecasting models by combining with these two frameworks.The experimental results show that the distributed short-term electric load forecasting model is better than the non-distributed short-term electric load forecasting model.Figure 61,table 3,reference 56...