Research on a Gas Concentration Prediction Algorithm Based on Stacking

Abstract

Machine learning algorithms play an important role in the detection of toxic, flammable and explosive gases, and they are extremely important for the study of mixed gas classification and concentration prediction methods. To solve the problem of low prediction accuracy of gas concentration regression prediction algorithms, a gas concentration prediction algorithm based on a stacking model is proposed in the current research. In this paper, the stochastic forest, extreme random regression tree and gradient boosting decision tree (GBDT) regression algorithms are selected as the base learning devices and use the stacking algorithm to take the output of each base learning device as input to train a new model to produce a final output. Through the stacking model, the grid search algorithm is studied to automatically optimize the parameters so that the performance of the entire system can reach the optimal parameters. Through experimental simulation, the gas concentration prediction algorithm based on stacking model has better prediction effect than other integrated frame algorithms and the accuracy of mixed gas concentration prediction is improved.

Keywords: automatic grid search algorithm; ensemble learning; model fusion; regression algorithm.

Figure 1

Schematic diagram of extreme random tree algorithm.

Figure 2

Stacking algorithm implementation process.

Figure 3

Stacking algorithm implementation form.

Figure 4

Model construction of stacking algorithm.

Figure 5

Random forest parameter sensitivity test results. (a) n_estimators parameter sensitivity test results. (b) max_features parameter sensitivity test results. (c) min_samples_split parameter sensitivity test results. (d) min_samples_leaf parameter sensitivity test results. (e) max_leaf_nodes parameter sensitivity test. (f) max_depth parameter sensitivity test.

Figure 6

TGS2600 sensor characteristic diagram.

Figure 7

Sensor response graph under Et_L_Me_H tag.

Figure 8

Comparison diagram of fitting algorithm. (a) Prediction result before stacking parameter optimization. (b) Prediction result after stacking parameter optimization.

Figure 9

Comparison diagram of fitting algorithms. (a) Stacking algorithm prediction result. (b) Blending algorithm prediction result. (c) Bagging algorithm prediction result. (d) Averaging algorithm prediction result. (e) SVR algorithm prediction result.