传递熵变量选择的非线性系统时序预测模型

Variables Selection for Nonlinear System Time Series Prediction Model by Transfer Entropy

对于复杂非线性系统,难以准确分析其中各变量之间相互影响关系,且其原始数据时序不统一。因此对于时序预测模型,上述情况会产生辅助变量冗余和时序混乱进而影响预测效果。提出了一种基于传递熵变量选择的非线性系统时序预测模型,利用因果关系进行辅助变量的筛选;检测变量之间的迟滞,据此进行时序统一。通过一个简单的非线性模型对算法进行验证,得到了和理论分析相吻合的结果,并利用不同模型分别验证了算法的有效性。最后,使用某600MW燃煤机组的实际数据对模型进行验证。实验结果证明:该预测模型与之前的方法相比,不依靠机理分析,通过较少的辅助变量得到准确的预测结果和更好的泛化能力,节约了运行时间,可以满足现场运行要求。

For complex nonlinear systems, it is difficult to accurately analyze the relationships between variables, and the sequence is not uniform. For the time series prediction model,the abovementioned situation will cause redundancy of secondary variable and the sequence chaos and then affect the prediction performance. A nonlinear system time series prediction model based on transfer entropy(TE) variable selection was proposed. The causal-effect relationship was used to select the secondary variables. The causal-effect between the variables was detected, and the sequence was unified accordingly. The method was validated by a simple nonlinear model, and the results were consistent with the theoretical analysis. The method was verified by different models as well.Finally, the model was validated by using real data from a600 MW coal-fired unit. Compares with the previous methods,the prediction model does not rely on mechanism analysis, and obtains accurate prediction results and has better generalization ability through fewer auxiliary variables, which saves running time and can meet the requirements of field operation.