ABSTRACT In this thesis, some aspects of the non-fixed neural network for power load forecasting are discussed. Unlike traditional fixed neural network technique, the structure of neural network is non-fixed during its training and testing phases. Based on the characteristic of the desired forecasting day, the number of input node utilized is changeable. The modified learning algorithms, including fuzzy back-propagation learning algorithm and stochastic back-propagation learning algorithm, will be used in the load forecasters we developed. For precise input selection of the neural network model, the analysis of mutual relationship between load and temperature and gray relational analysis between desired forecasting load and the related previous load are studied. Two types of load forecasting, i.e., peak load forecasting and hourly load forecasting, are investigated. Short term (one-to-several-day-ahead) load forecasting is considered in this research. Hourly loads and relevant temperature data from 1992 to 1998 provided by Taipower Utility and the Central Weather Bureau is implemented for this research. For demonstrating the feasibility and superiority of the forecasters we develop, several forecasting models, including fixed neural network with constant learning rate and momentum, recursive time series model, and artificial neural network short term load forecaster (ANNSTLF) proposed by [Kho.2], are also performed for a comparison. From the results of the simulation, better performances could be obtained by the methods we proposed. Not only the over-training phenomenon is obviously reduced, the forecasting accuracy and the learning speed of the neural model are also effectively improved.