Study Of Electricity Price And Load Forecasting Model Based On Multi-task Deep Learning

With the developments of UPIo T（Ubiquitous Power Internet of Things）and distributed energy,current electricity price and power load forecasting technology based on big data is facing the new challenge.However,most of the current researches regard electricity price prediction and power load prediction as two independent problems to build the corresponding models which do not consider the correlation between electricity price and power load.Therefore,these models always had low prediction accuracy.In this thesis,the correlation between electricity price and power load is considered,and a multi-task deep learning model is proposed which can simultaneously forecast the electricity price and the power load.The main contents of this thesis include:（1）The correlation,autocorrelation and seasonality of electricity price and power load time series are analysed.The correlation analysis shows that there is a moderate intensity correlation between electricity price and power load time series,and the correlation coefficient is 0.5;The autocorrelation analysis shows that when the time delay is 24 hours,the autocorrelation between electricity price and power load reaches the peak value and has a weekly variation trend;Seasonal analysis shows that electricity price and power load have seasonal variations.Thus,electricity price and power load time series are non-stationary time series.These analysis results provide guidance for the following prediction model establishment.（2）Because the time series of electricity price and power load are non-stationary.Therefore,the Adaptive Variational Mode Decomposition（AVMD）algorithm based on Particle Swarm Optimization（PSO）is proposed in this thesis.AVMD is used to stabilize the electricity price time series and power load time series.This algorithm can search the fidelity constraint parameter values and decomposition model numbervalues adaptively,and achieve efficient time series decomposition at the same time.（3）This thesis has proposed two kinds of electricity price and power load forecasting models based on multi-task deep learning.The first model takes to load as the main task and electricity price as the auxiliary task.The second model takes to electricity price as the main task and load as the auxiliary task.These models skillfully combined the characteristics of two-dimensional Convolutional Neural Network（CNN）,Temporal Convolutional Network（TCN）and multi-task deep learning framework,and realized the high-precision simultaneous prediction of electricity price and power load.（4）Finally,the proposed models are applied to the New York City data set which are selected from New York Independent System Operator（NYISO）.Comparing five models,such as LSTM,TCN,CNN^1D-TCN,CNN^2D-TCN and the proposed,the results show that the proposed models have the best performances.It is also verified that the multi-task deep learning model’s accuracy is better than that of the single-task deep learning model in load and price forecasting.