Yongxing Cao

Power Load Forecasting Based on A Hybrid Optimum Training Algorithm Embedded with Chaos Sequence

ANN using BP is widely used in power load forecasting. But there are some existed problem of the BP algorithm: (1) Convergence speed is slow, usually convergence needs more than one thousand times; (2) Objective function is prone to getting into local minimum.. How to overcome the shortcoming that convergence speed is slow and network is prone to trapping in local minimum has not been resolved. Training algorithm is put forward in the paper, which is based on three adjustable parameters activation function. Although BP-Adjustable activation function algorithm raised convergence speed of neural network, the essence of algorithm is to seek the most optimal value following the direction of degressive grads. When initial value is confirmed, the decent course is also confirmed. So the problem of trending local minimum also exists, and it has great contact with confirming initial value. The paper says that BPAA algorithm has the good ability of searching in partial area and ability of searching in global area. So, it can combine with two algorithms to make the best of LOA (Logistic optimal algorithm)s fully searching ability and BP-AA algorithms partial searching ability. Using BP-AA algorithm to resolve the weight value and parameters in neutral network. When getting in partial least, using LOA algorithm can choose new initial value of every parameter to jump out partial least. Therefore, the paper raised the Adjustable activation function and grad optimism training algorithm embedded Logistic chaotic mapping in BP networks training. This algorithm is not only efficient but also difficult to get in local minimum. The algorithm can converge to fully optimum with probability of 1. it is called BP-AAEC (Adjustable Activation function and Embedding Chaos algorithm) Algorithm. Then, combined with randomness and ergodic property of chaos, the hybrid training algorithm embedded dual searching of chaos mapping is brought forward. Capability testing and experiment has appro- - ved that the improved algorithm can achieve the requirement.