摘要
以某地区购网有功功率的负荷数据为背景,建立了三个BP神经网络负荷预测模型——SDBP、LMBP及BRBP模型进行短期负荷预测工作,并对其结果进行比较。针对传统的BP算法具有训练速度慢,易陷入局部最小点的缺点,采用具有较快收敛速度及稳定性的L-M优化算法进行预测,使平均相对误差有了很大改善,具有良好的应用前景。而采用贝叶斯正则化算法可以解决网络过度拟合,提高网络的推广能力,使平均相对误差和每日峰值相对误差降低,但收敛速度过慢(慢于SDBP模型),不适于在实际应用中采用。
Based on the load data of meritorious power of the power system of some area, three BP ANN models, named SDBP model, LMBP model and BRBP Model, are established to carry out the short-term load forecasting work, and the results are compared. As for the traditional BP algorithm has some unavoidable disadvantages such as the slow training speed and the feasibility of being plunged into local minimums, an optimized L-M algorithm, which has a quicker training speed and better stability, should be applied to forecast, which can effectively reduce the mean relative error. So it has a good application prospect. In contrast, Bayesian Regularization can overcome the excessive fitting, improve the generalization of an ANN and reduce the mean relative error and the relative error of everyday peak values, it is not suitable for actual application because of its slow training speed （slower than SDBP model）.
出处
《继电器》
CSCD
北大核心
2007年第6期49-53,共5页
Relay
关键词
短期负荷预测
人工神经网络
L-M算法
贝叶斯正则化算法
优化算法
short-term load forecasting （STLF）
ANN
Levenberg-Marquardt algorithm
Bayesian regularization
optimized algorithms
