Abstract: In the actual process of oilfield development, affected by the production of new blocks, the adjustment of development plans and the "t three mining" measures, the annual output data will show multi-peak form. As the classical Hubbert, HCZ and other models cannot directly fit the multi-peak data sequence, the multi-peak prediction model of oilfield production based on machine learning is studied. Based on the Hubbert model, the piecewise least squares fitting is performed for multi-peak data sequence, the penalty term controlling the number of segments is introduced into the fitting error function. Using dynamic programming algorithm, the multi-peak Hubbert prediction model for the optimal segment is automatically obtained. The model is applied to the actual oilfield production data, and the prediction results achieve the expected purpose. This paper presents a method to build a multi-peak prediction model of oilfield production through automatic optimal segmental linear regression learning. In practical application, it has the advantages of simple modeling and strong adaptability.

Key words: oil production forecasts, machine learning, dynamic programming, multiple peaks, Hubbert model