Short term load forecasting is performed by data, whose validation is subjected to measurement systems errors, transmission errors and …, in addition unpredictable events and load shedding, voltage collapse lead to inappropriate load data and unusual load profiles. In this thesis the significant of the refinement of improper load data and different schemes for omitting improper load profiles is investigated. In addition a scheme based on absolute normalized residual for modification of improper hour load data instead of omitting is proposed, and for Isfahan power system a load forecasting based on these algorithm is performed. And the performance of this algorithm in decreasing the STLF errors is assessed. Finally the impact of omitting and modifying of improper load data in decreasing the STLF errors in schemes based on perceptron neural networks is investigated. Because of special load pattern and lock of similar load profiles, load forecasting in holidays has been always one of the challenging problems of STLF systems. In addition various holidays and displacement of some of them and the existence of inappropriate load data lead to increasing STLF errors. In this thesis forecasting the holidays based on the Absolute Normalized Residual is performed. The results show that the load data modification in the algorithms can significantly decrease the errors of forecasting in normal days and specially holidays. Keywords Load Forecasting, Improper Load Data, Absolute Normalized Residual