The mud pit volume (MPV) model is of great importance in evaluating the bottom hole pressure (BHP). In this paper, an online hybrid model is developed to predict MPV considering the drilling characteristics of data pollution, multi-variable, strong nonlinearity, and time series characteristics. First, the mutual information and fast Fourier transform method are introduced to filter data noises and determine the model inputs. Then, back propagation neural network (BPNN) method and support vector regression (SVR) method are used to establish the submodels, and the submodels are combined based on three evaluation criteria. After that, the combination model is fine-tuned according to the time series trends of MPV based on the long short-term memory neural network (LSTMNN). Finally, a modified sliding window method is developed to update the hybrid model constructed by SVR, BPNN and LSTMNN. The simulation results based on actual drilling data show that the online hybrid model has higher accuracy than other prediction models, and the online hybrid model can follow the time series characteristics of MPV, which validates the effectiveness of the developed model.

泥浆体积（MPV）模型对于评估井底压力（BHP）至关重要。在本文中，考虑到数据污染的钻探特征，多变量，强非线性和时间序列特征，开发了一种在线混合模型来预测MPV。首先，引入互信息和快速傅里叶变换方法来过滤数据噪声并确定模型输入。然后，采用反向传播神经网络（BPNN）方法和支持向量回归（SVR）方法建立子模型，并根据三个评估标准对子模型进行组合。之后，基于长短期记忆神经网络（LSTMNN），根据MPV的时间序列趋势对组合模型进行微调。最后，开发了一种改进的滑动窗口方法来更新由SVR，BPNN和LSTMNN构建的混合模型。基于实际钻探数据的仿真结果表明，该在线混合模型具有比其他预测模型更高的准确性，并且该在线混合模型可以遵循MPV的时间序列特征，验证了所开发模型的有效性。