Selecting the reference points of current phase point and gaining higher forecasting accuracy of the peak-valley is important for short-term load forecasting based on phase space reconstruction. The daily peak-valley load time series

which is proved chaotic by fractal dimension and Lyapunov exponent analysis

is constructed based on the daily load curves. The load per hour is then corrected through peak-valley load prediction by the phase space reconstruction. An effective method composed of rough search and fine search is presented to choose reference points with the purpose of improving the forecasting precision of chaotic time series．The rough search is used to search some neighboring points according to the number of reference points required by forecasting model

the false neighboring points are kicked off through fine search in terms of the time evolution relativity

and the time weights is also introduced to consider the time sequence of different coordinates in the phase space. The simulation results of practical load forecasting show that the proposed method to select reference points and peak-valley load correction are more effective.