基于LSTM-Adaboost的多晶硅生产的能耗预测

ENERGY CONSUMPTION PREDICTION OF POLYSILICON PRODUCTION BASED ON LSTM-ADABOOST

在多晶硅的生产过程中,还原工序是最重要的工序,其能耗占综合能耗的60%～70%。针对还原工序能耗影响因素复杂,非线性,传统预测方法预测精度低等问题,提出基于LSTM-Adaboost循环神经网络多晶硅生产过程的能耗预测模型。通过PCA算法对多晶硅生产过程的能耗影响因素按贡献率提取主成分;采用正则化方法优化LSTM的目标函数并引入Adaboost算法对LSTM模型优化;构建LSTM-Adaboost预测模型,实现能耗预测。实验结果表明,相比于LSTM模型和BP模型,LSTM-Adaboost模型预测精度更高。以某企业多晶硅还原工序为例,验证该能耗预测的有效性。

In the production of polysilicon, the reduction process is the most important process, and its energy consumption accounts for 60%~70% of the total energy consumption. The reduction process has complex influence factors in the energy consumption, and is non-linear. The traditional prediction methods have low accuracy. To solve these problems, we proposed the energy consumption prediction model for the polysilicon production process based on LSTM-Adaboost recurrent neural network. PCA algorithm was used to extract principal components of influencing factors on energy consumption in the polysilicon production process according to the contribution rate. The regularization method was adopted to optimize the objective function of LSTM and the Adaboost algorithm was introduced to optimize LSTM model. The LSTM-Adaboost prediction model was constructed to realize energy consumption prediction. The experimental results show that compared with LSTM model and BP model, LSTM-Adaboost model has better prediction accuracy. Taking the polysilicon reduction process of an enterprise as an example, the validity of the energy consumption prediction is verified.