PSO演化神经网络集成的边际电价预测新方法

为了克服神经网络模型结构和参数难以设置,学习算法收敛速度慢等缺点,提出了一种基于粒子群优化的演化神经网络集成新模型对日前交易电力市场的边际电价进行预测。该模型将边际电价预测问题转化为神经网络实际输出与预测输出误差最小化问题,首先采用粒子群优化算法把神经网络的结构和权重映射成问题空间中的粒子,通过粒子速度和位置更新方程进行粗学习,获得多个相对占优的神经网络结构和初始权重并构成神经网络集成预测模型,然后采用梯度学习算法和交叉验证对神经网络集成单元的权重进行细学习,并以误差最小的神经网络集成单元的输出作为神经网络集成预测模型的输出。运用此方法对加州日前交易电力市场的边际电价进行了日预测,结果表明其优于三层BP神经网络预测方法。

Novel Approach of Market Clearing Price Forecasting: PSO Based Evolutional Neural Network Model

A novel evolutionary neural network ensemble model based on particle swarm optimization(PSO) is proposed to forecast market clearing price(MCP) in day-ahead electricity market.In the proposed model,MCP forecasting problem is converted into error minimization problem between actual output and desired output of neural network.At first,construction and weights in neural network are initiated to be particles in problem space,and then neural network are extensively trained by both velocity update equation and position update equation which are defined by particle swarm optimization algorithm.As a result,a number of neural networks with relatively predominated construction and weights over other neural networks are obtained,and neural network ensemble is constructed by combination of neural network units with better constructions and better initial weights.Next,weights in neural network units are intensively trained by gradient learning algorithm and cross validation.The output of the neural network unit with minimal error is regarded as the output of neural network ensemble.The novel model is applied to forecast MCP of California day-ahead electricity market.Day-forecasting results show that the model can obtain satisfactory forecasting accuracy and is superior to three-layer BP neural network.