There are many factors which can affect the electricity prices, but in fact, we can not get all of the information of these factors. In this state, to improve the predictive precision, this paper analyzes the chaotic property of the price and load time series and reconstructs the attractors using C-C theory. Besides, the similarity search technique in data mining is adopted to find the most similar time series as training date. Then BP neural network which has high ability to self learning and self adapting is used to find the reconstructive function. The actual data of American PJM power market is forecasted based on the theory above. The results have good predictive precision, especially in rest days.