Electricity Power System Energy Forecasting Research Based On Loss Functions And Neural Networks

As an important part of the stable operation of power system,the energy forecasting of electric power system has important research significance.Wind energy is one of the main sustainable energy on the supply side of power system,and accurate wind energy forecasting is an important part of the safe,efficient and stable operation of wind farms.Power load forecasting is the main part of the energy forecasting of power system in demand side.Accurate load forecasting is important for power companies to formulate rational planning of power distribution.Current electric power system energy forecasting methods exist poor robustness,complex random noise,difficulty in extracting linear and nonlinear features,limitations in practical applications and poor long-term feature extraction.This thesis proposes the following research contents to solve above problems:(1)A robust penalized extreme learning machine model based on robust loss function and unified penalized regression framework is designed.This model can solve the problems that the extreme learning machine is sensitive to outliers and the model structure is difficult to be determined.Compared with the extreme learning machine,the objective function of the robust penalized extreme learning machine model can improve its robustness by replacing the square loss function with exponential square loss and lncosh loss.The exponential square loss and lncosh loss have better robustness.In addition,the model structure can be punished by adding penalty terms(Ridge penalty and Lasso penalty)in the unified penalty regression framework.Therefore,the model structure can be adaptively determined.We apply the proposed model to the experiments of wind energy data set in northern China.The use of Lasso penalty shows the excellent ability to compress the hidden layer nodes of the extreme learning machine and improve the prediction accuracy.The use of Ridge penalty can significantly improve for the prediction performance of the model.(2)A robust auto-regression bidirectional GRU short-term forecasting model based on robust loss function and neural network is designed.This model can solve the problems that linear feature and non-linear feature extracting difficultly in wind energy forecasting and load forecasting.The auto-regression algorithm and bidirectional GRU neural network are introduced for capture linear and non-linear feature respectively.The self-attention mechanism is applied to extract the correlation features of hidden features encoded by bidirectional GRU neural network at different timestamps.For the problem of outliers in wind energy data sets and load data sets,the adaptive re-scaled Huber loss function is introduced to train the neural network.The hyper-parameters meeting the data distributions in this loss can be iteratively estimated by optimizing the ‘working’likelihood function to improve the robustness and fitting ability to complex random noise.Experiments on wind energy and Australian load data sets verify the better forecasting performance of the proposed model.(3)A multi-scale decomposition neural network short-term forecasting model is designed.This model can solve the problems of repeated decomposition and modeling of the traditional load forecasting methods based on signal decomposition methods.The Fourier transform of the load series is approximately realized based on a deep neural network.This Fourier transform can adaptively decompose the load series into periodic components with different frequencies and amplitudes,and avoid the repeated process of decomposition and modeling.For the problem of outliers and random noise in load data sets,the adaptive re-scaled lncosh loss function is applied as the objective function based on the robust regression.The ‘working’ likelihood function of the adaptive re-scaled lncosh loss is derived to iteratively optimize the hyper-parameter conforming to the data distribution of random noise,so as to improve the robustness and the fitting ability of complex random noise.Experiments on load data sets in Portugal and Australia verify the better forecasting performance of the proposed model.(4)A novel improved Auto-correlation mechanism is designed and a multi-granularity Autoformer point and probabilistic interval forecasting model is proposed.The multi-granularity Autoformer can solve the problem of difficult long-term feature extraction in Autoformer.The improved Auto-correlation mechanism introduces the shared Q-K mechanism to map the query matrix Q and the key matrix K of the self-attention mechanism into the same feature space to calculate the correlation between different timestamps more appropriately.This mechanism also extends the Auto-correlation mechanism from single sample to multiple samples,which can extract the correlation features with different granularity.In addition,the Autoformer is extended to point and probabilistic interval prediction model based on quantile loss function.Experiments on load data sets in Portugal,the United States,ISO New England and Australia verify the better prediction performance of the multi-granularity Autoformer in point and probabilistic interval prediction.