Modified GMDH networks for oilfield production prediction

Abstract

ABSTRACTThe self-organizing Group Method of Data Handling (GMDH) functional network is effective in predicting oilfield production. During operation the division of data sample depending on artificial classification cannot lead to global optimum in great probability and the variables are probably eliminated early in the iterative process in traditional GMDH algorithm. Recent years, GMDH model has been improved through many artificial intelligent models, but few people take the optimization of the model structure into account. In this paper, different training and testing set grouping and the effects of variables transmission were studied. The modified GMDH algorithm was optimized using the original variables preservation method and the random sample method, which was applied to the oilfield production forecasting simulation. The results of the modified GMDH algorithm, the traditional GMDH algorithm, ANNs and the empirical equations for predicting annual oil production were compared. The simulative results indicated that the modified GMDH model was the best tool for data-fitting with lowest error (RMSE = 13.9440, MAPE = 0.1121 and SI = 0.0378) and highest accuracy (R = 0.9984).