The actual generation price does not always satisfy the normal distribution， so it is necessary to carry on further studies on the power purchasing strategies for power supply companies at multi-markets and multi-stages. At first， the characteristic parameters of the fractal distribution of actual generation price in American PJM power market are estimated， and its fitting performance is checked by Kolmogorov-Smirnove（K-S） method. Then the load is estimated by average reverse model（AR（1））， and the efficiency of this model is verified by the actual load of PJM power market. Finally， given that generation price satisfies fractal distribution， formulae for evaluating power purchasing risks for power supply companies are derived. Based on the above， the power purchasing model for power supply company is constructed based on investment portfolio theory， and the model is transformed to the linear programming model by equivalence theory. Case study shows that the constructed model can truly reflect the risk exposed to the power supply company.