A Time Series Prediction Method Based on Self-Adaptive RBF Neural Network

摘要

Time Series Prediction is widely used in our daily life. We propose a forecasting method based on RBF neural network for time series prediction in this paper. This approach consists of two phases, training phase and working phase. During training phase, we integrate subtractive clustering method and k-means method to generate the centers of RBF neural network, which can cover the shortage of only using k-means method. Then we use orthogonal least squares method to calculate the weight for the output layer. And during the working phase, we bring in a performance evaluation mechanism to determine whether to update the training set or not. If the output data of the network do not perform well, then we put the relative input data into training set and go back to the training phase to reconstruct the network. The experiment shows that this approach improves the prediction accuracy than the traditional method only using k-means to train the network, and it makes the RBF neural network has the ability to change with different input data.