基于GAD-BP神经网络的短期负荷预测

Short-term load forecasting based on GAD-BP neural network

短期负荷的精准预测可以提高配网运行的安全性、可靠性、经济性,虽然BP神经网络在数据训练和关联记忆方面具有良好应用效果,但其在短期电负荷预测领域存在局部最优解和收敛速度慢的不足。因此,提出了一种改进的BP神经网络算法,首先构建一个全局化加速BP(globally accelerated BP,GAD-BP)神经网络模型,在模型参数预训练过程中,采用遗传算法全局寻优,通过LM算法优化初始权值。其次结合Adagrad方法逐参数除以迭代历史梯度平方根和来动态调节学习率,使其更快收敛于最优解。最后将GAD-BP与St-BP、LSTM进行仿真对比,结果表明,GAD-BP神经网络模型计算时间短,且预测精度达到了0.981 2×10-3。

The accurate prediction of short-term load is crucial to improve the safety, reliability and economy of the operation of distribution network. Although BP neural network has positive application effect in data training and associated memory, it is deficient in global optimization and convergence speed in the application of short-term electrical load prediction. Therefore, an improved BP neural network algorithm is proposed. Firstly, we build a the globally accelerated BP neural network model.In the pre-training process of model parameters, global optimization is adopted by genetic algorithm. The initial weights are optimized by LM algorithm. Secondly, the learning rate is dynamically adjusted by the Adagrad.Each parameter is divided by the iterated history gradient square root sum to make it converge to the optimal solution faster.Finally, the GAD-BP was compared with St-BP and LSTM.The results show that the GAD-BP neural network model has short computation time.The prediction accuracy also reaches 0.981 2×10-3.