Research On Short-term Power Load Forecasting Based On Deep Neural Network

Load forecasting is an important part of smart grid construction,energy management,and sustainable design of power systems.It greatly impacts reliable grid operation,facility planning,and other decisions.In the new context of the energy internet,the development of new energy technologies and energy policy changes have increased uncertainty in grid planning.Many renewable energy sources and flexible loads can affect grid operations caused by variable weather and variable electricity prices,which introduce additional uncertainties in load forecasting.These uncertainties prevent load forecasting algorithms from capturing the interdependencies between Multivariate Time Series(MTS),including load,temperature,humidity,and electricity prices.In addition to these uncertainties,it is difficult for existing load forecasting models to capture the nonlinear,non-stationary,or unstable characteristics of the MTS generated by the forecast dependencies,including temperature,humidity,load,and electricity price.In recent years,deep learning-based frameworks have shown superior forecasting performance.Load forecasting usually involves a variety of complex nonlinear factors,such as weather variations and holidays,which are difficult to be handled effectively by traditional linear models,while deep learning models can model these factors from a nonlinear perspective and can capture dynamic trends at different time scales.In this thesis,using the basic theory of load forecasting as a guide and deep learning methods as a tool,the problem of short-term load forecasting based on deep neural networks is studied.The mainstream deep neural network-based prediction methods are summarized,including Long Short Term Memory Network(LSTM),Convolutional Neural Network(CNN),the attention mechanism,and the Temporal Convolutional Network(TCN)and Transformer model proposed in recent years,and the working principles and characteristics of each model are analyzed.The main research contents of this thesis are as follows.(1)Using two real-case electricity load datasets,this thesis investigates the influence of electricity load-related factors(including weather conditions,dates,electricity prices,and other factors)on the load.To avoid redundancy of model input data and increase the risk of overfitting the model.The design of the prediction model input and output variables was completed by analyzing the influencing factors and determining the feature selection of the dataset.Pre-processing of the experimental dataset was performed,including missing value filling,outlier processing,and data normalization.(2)To address the problem that the complex and variable interdependencies among MTS are difficult to capture and the interference brought by uncertainty to the prediction.In this thesis,a model based on TCN networks combined with attention mechanism for short-term load forecasting is proposed.The proposed model uses the TCN network as a feature extractor for load data,constructs very long effective histories by dilated causal convolution,and the dilated causal convolution has flexible sense field sizes for modeling long-term dependencies of load MTSs.After the TCN network,the attention mechanism suggested on the LSTM pairing is used to capture the important information of the input sequence by adaptively weighting the input MTS to reduce the interference of uncertainties on the prediction.An Australian load forecasting dataset is studied as a case study to complete the analysis work to illustrate the effectiveness of the model.(3)To further improve the accuracy of short-term load forecasting,a Time Window Attention(TWA)-based short-term load forecasting model is proposed by stacking Self-attention(SAT)-LSTM learners in this thesis.Each base learner adopts an encoderdecoder structure,where the encoder is a Self-attention mechanism paired with a 1DCNN,and the Self-attention mechanism has the ability to draw global dependencies between inputs and outputs,which is used to capture the long-term dependencies of the input MTS data and extract the input features.A TWA algorithm is proposed for the decoder,which captures the strong spatio-temporal dependencies of the MTS in multiple time steps within a fixed-length time window and adaptively weights the LSTM hidden states to reduce the impact of uncertainty on load prediction.In the prediction model section,stacking of multiple base learners is performed to increase the model robustness,and fusing multiple base learners enables the model to increase the covered feature space by jointly focusing on various subspaces to improve the prediction accuracy.The proposed prediction model is experimented on two different types of datasets for prediction,illustrating the effectiveness and advantages of the proposed model.In addition,experiments with synthetic datasets and ablation models were constructed to further validate the proposed model’s ability to capture interdependencies and mitigate the effects of uncertainty.The experimental results show that the stacked SAT-LSTM structure with TWA improves the accuracy of load prediction.