Abstract: When soft sensor model is used to predict dynamic fluid levels in oil production， it will gradually degenerate during the process， resulting in the larger deviation of prediction results and the difficulties to be used in practical oilfield production. To solve this problem， a new just-in-time model based on the similarity of subspaces was proposed to realize adaptive dynamic updates for a prediction model of dynamic fluid level. According to the production data， the similarity of subspaces was calculated to improve the accuracy of selecting modeling samples. Two memory parameters were designed to change the update method in traditional just-in-time learning model， which could reduce the amount of calculation and improve the prediction accuracy of dynamic fluid level. Compared with the traditional just-in-time learning algorithm， the improved method has better measurement accuracy and adaptation for the prediction of dynamic fluid levels. The example showed that the proposed method was fitted in with the standard of oil production and could be applied to actual production.

Key words: subspace, just-in-time learning, model update, similarity, dynamic fluid level