针对甘肃高比例新能源双边现货市场价格数据的非线性特征以及主流损失函数的缺陷,本文主要研究如何使用Huber损失函数与神经网络精确模拟日前电力现货价格的时间序列,旨在系统解决甘肃日前电力现货价格预测时反向传播损失函数的选择。...针对甘肃高比例新能源双边现货市场价格数据的非线性特征以及主流损失函数的缺陷,本文主要研究如何使用Huber损失函数与神经网络精确模拟日前电力现货价格的时间序列,旨在系统解决甘肃日前电力现货价格预测时反向传播损失函数的选择。首先分析了甘肃电力现货市场的运行特点及问题,建立了其电力现货价格的神经网络模型;为了模型的优化建立了反向传播算法中损失函数应该满足的期望属性,并筛选出满足属性的Huber损失函数用于甘肃电力市场现货价格预测,还对市场因素的相关性进行了分析;通过实证研究,在MSE、MAE和Huber上训练了RNN神经网络,发现利用2024年1月1日至2024年6月30日期间甘肃电力现货价格数据在Huber上训练的模型比在MSE和MAE上训练的模型提供了更准确的日前电价预测,Huber函数是训练神经网络预测甘肃日前电力现货价格的最佳选择。而且Huber函数的回溯结果和稳定性比MSE和MAE都好,即利用Huber损失函数预测数据的精准性和稳定性更强。In response to the nonlinear characteristics of the bilateral spot market price data for high-proportion new energy in Gansu and the shortcomings of mainstream loss functions, this paper primarily investigates how to use the Huber loss function with neural networks to accurately model the time series of day-ahead electricity spot prices, aiming to systematically address the selection of the backpropagation loss function for day-ahead electricity spot price forecasting in Gansu. Firstly, the operational characteristics and issues of the Gansu electricity spot market are analyzed, and a neural network model for its electricity spot price is established;to optimize the model, the expected properties that the loss function should meet in the backpropagation algorithm are established, and the Huber loss function, which meets the properties, is selected for forecasting the spot price of the Gansu electricity market, and the correlation of market factors is analyzed;through empirical research, Recurrent neural networks (RNN) are trained on MSE, MAE, and Huber, and it is found that using the Gansu electricity spot price data from January 1, 2024, to June 30, 2024, the model trained on Huber provides more accurate day-ahead electricity price forecasts than the models trained on MSE and MAE. The Huber function is the best choice for training neural networks to predict day-ahead electricity spot prices in Gansu. Moreover, the retrospective results and stability of the Huber function are better than those of MSE and MAE;that is, the precision and stability of predicting data using the Huber loss function are stronger.展开更多
文摘针对甘肃高比例新能源双边现货市场价格数据的非线性特征以及主流损失函数的缺陷,本文主要研究如何使用Huber损失函数与神经网络精确模拟日前电力现货价格的时间序列,旨在系统解决甘肃日前电力现货价格预测时反向传播损失函数的选择。首先分析了甘肃电力现货市场的运行特点及问题,建立了其电力现货价格的神经网络模型;为了模型的优化建立了反向传播算法中损失函数应该满足的期望属性,并筛选出满足属性的Huber损失函数用于甘肃电力市场现货价格预测,还对市场因素的相关性进行了分析;通过实证研究,在MSE、MAE和Huber上训练了RNN神经网络,发现利用2024年1月1日至2024年6月30日期间甘肃电力现货价格数据在Huber上训练的模型比在MSE和MAE上训练的模型提供了更准确的日前电价预测,Huber函数是训练神经网络预测甘肃日前电力现货价格的最佳选择。而且Huber函数的回溯结果和稳定性比MSE和MAE都好,即利用Huber损失函数预测数据的精准性和稳定性更强。In response to the nonlinear characteristics of the bilateral spot market price data for high-proportion new energy in Gansu and the shortcomings of mainstream loss functions, this paper primarily investigates how to use the Huber loss function with neural networks to accurately model the time series of day-ahead electricity spot prices, aiming to systematically address the selection of the backpropagation loss function for day-ahead electricity spot price forecasting in Gansu. Firstly, the operational characteristics and issues of the Gansu electricity spot market are analyzed, and a neural network model for its electricity spot price is established;to optimize the model, the expected properties that the loss function should meet in the backpropagation algorithm are established, and the Huber loss function, which meets the properties, is selected for forecasting the spot price of the Gansu electricity market, and the correlation of market factors is analyzed;through empirical research, Recurrent neural networks (RNN) are trained on MSE, MAE, and Huber, and it is found that using the Gansu electricity spot price data from January 1, 2024, to June 30, 2024, the model trained on Huber provides more accurate day-ahead electricity price forecasts than the models trained on MSE and MAE. The Huber function is the best choice for training neural networks to predict day-ahead electricity spot prices in Gansu. Moreover, the retrospective results and stability of the Huber function are better than those of MSE and MAE;that is, the precision and stability of predicting data using the Huber loss function are stronger.
