Predicting the construction cost of high standard farmland irrigation projects using NGO-CNN-SVM

Construction cost is difficult to predict in high-standard farmland projects, due to the short construction period. In this study, the construction cost was predicted for the single-unit irrigation projects, in order to improve the prediction accuracy. The key influencing factors were then screened from the construction cost of high-standard farmland irrigation projects using random forest (RF). Then, convolutional neural networks (CNN) and support vector machine (SVM) were combined to construct a CNN-SVM-based prediction model, in order to improve the prediction accuracy over a single model. The penalty parameter C and kernel function parameter g of the CNN-SVM model were optimized by northern goshawk optimization (NGO). The NGO shared the higher convergence speed and stronger optimization, compared with the rest. Finally, the prediction model (NGO-CNN-SVM) was obtained for the construction cost of high-standard farmland irrigation projects. The data was collected from the irrigation project in the high-standard farmland in Liaoning Province from 2018 to 2023. Coefficient of determination (R²), mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) were taken as the accuracy indexes for analysis. The results show that the NGO-CNN-SVM model of construction cost in the channel project shared the MAE and RMSE lower than 0.615 and 0.512 million yuan, respectively, where the R² reached more than 0.968, and the relative error was less than 4.21%; In the project of the inlet sluice gate, the MAE and RMSE were lower than 0.610 and 0.536 million yuan, respectively, where the R² reached more than 0.966, and the relative error was less than 4.41%; In the project of bridge and culvert, the MAE and RMSE were less than 0.494 and 0.477 million yuan, where the R² reached more than 0.970, and the relative error was less than 3.548%. Taking channel engineering as an example, the deep learning network model (CNN) was reduced by about 34%, 20%, 33% in the MAE, RMSE, MAPE, and improved by 3% for the R², respectively, compared with the traditional back propagation neural networks. A higher efficiency of prediction was achieved in the less complexity and nonlinear data dimension. In addition, the CNN-SVM model was significantly reduced by about 40%, 46%, 41% in the MAE, RMSE, MAPE, and improved by 4% for the R², respectively, compared with the model CNN, indicating better performance than that of the single model. Compared with the CNN-SVM model, the MAE, RMSE, MAPE of the NGO-CNN-SVM model were reduced by about 22%, 25%, 17%, and improved by 4% for the R², respectively, indicating that the randomly generated hyperparameters led to the low generalization and prediction accuracy. The optimal optimization of the hyperparameters had further improved the performance of the improved model. In summary, the feature selection, model fusion, optimization, and comparison of different models show that the NGO-CNN-SVM model shared higher prediction accuracy and generalization. The model can be applied to directly establish the index system, in order to estimate high-standard farmland projects. The finding can provide the theoretical basis and support to formulate the construction cost and budget quota of irrigation projects in high-standard farmland.