Research on Settlement Prediction of Small Water Conservancy Project based on ELM Model Optimized by Genetic Algorithm

To find suitable for small water conservancy engineering standard method for prediction of subsidence. This paper based on the genetic algorithm GA optimization extreme learning machine, three different ELM model activation function. From this, six computational models are obtained. According to the input of groundwater dynamic changes, precipitation, temperature and soil four indicators of the two kinds of input combinations, a total of 12 kinds of model input. It’s concluded that the optimal settlement prediction model, the results showed that: Ga-ELMsin model shows high accuracy, and genetic algorithm can improve the calculation accuracy of ELM model. Groundwater dynamics is the main factor affecting settlement.


Introduction
With the growth of the domestic population and the rapid development of the economy, the requirements for the quality of the project are also gradually improved in the limited land resources. It is the focus of relevant industries to promote engineering development and improve residents' quality of life [1][2] . Small water conservancy projects are closely related to the life of residents. China's water conservancy industry is mainly concentrated in the construction of small water conservancy projects, which have direct and obvious benefits and are directly related to residents' daily life and water use for crops, and are of great significance to the improvement of residents' life [3][4] . The quality and operation management of small water conservancy projects may directly affect the daily work and life of residents. However, due to the small water conservancy projects' less investment and remote location, the implementation degree of project quality monitoring is low, which seriously affects their service life.
Settlement is one of the main reasons that affect the service life and durability of a project. The settlement is mainly affected by natural factors and human factors, resulting in deformation in the area where the project is located and resulting in certain damage to the main structure of the project due to additional stress [5] .Longterm settlement observation shows that there is a certain distribution law in the process from the construction period to the stable period, but the settlement is not easy to be discovered and has a long periodicity, which has become an important geological disaster and has been listed as the focus of prevention and observation [6] .
With the increasingly serious influence of settlement, it has become the research of relevant departments to find out the reasonable settlement prediction method while observing the settlement daily. Among them, the machine learning neural network model is widely used. Wang Lu and Gui Zhanfei [7] established a building settlement prediction model based on GM-ARMA-BP model, and pointed out that this combined model overcame the disadvantages of the single model. By integrating the advantages of each model, the prediction accuracy of the comprehensive model was improved by at least 50% compared with that of the single model. Feng Shaoquan et al. [8] established a prediction model for the settlement amount of elevated bridge piers based on the combined model of GA-BP-MC, and proved the scientific nature of the model based on the index of absolute error and relative error. Song Chuping [9] also established the settlement prediction model of deep foundation pit based on the improved BP neural network model, and pointed out that the improved BP neural network model had faster convergence speed and higher generalization ability. Peng Yuan [10] used genetic algorithm to optimize BP neural network model to predict the settlement of buildings, and achieved good results.
Extreme learning machine (ELM) is a new neural network algorithm. This algorithm overcomes the shortcoming of slow convergence speed of traditional BP neural network model and is applied to data training and prediction. The genetic algorithm facilitates the model to find the optimal path and find out the rules of training data. In this paper, a settlement prediction model for small water conservancy projects is established based on the ultimate learning machine (GA-ELM) algorithm of genetic optimization. mainly composed of input layer, hidden layer and output layer. Its basic structure is shown in Figure 1.
Let the activation function of hidden layer neurons be g (ω, X, b), then the expression of output layer of ELM model is: Where β j -the connection value between the hidden layer and the output layer; ω ij -the connection value between the input layer and the hidden layer; X i -Input layer variable value; b j -threshold for the JTH hidden layer.
The activation function of ELM model can be divided into three types: Sine function, Radbas function and Hardlim function. In this paper, three activation functions are respectively used to establish the settlement prediction model for small water conservancy projects. The specific formulas of the three functions are as follows: Sine function ELM sin : Radbas function ELM rad : Hardlim function ELM hard :

Genetic algorithm optimization of extreme learning machine model (GA-ELM)
The application of the principle of genetic algorithm in the optimization of neural network learning can help the neural network model find the optimal solution as soon as possible and shorten the training time of the model. The specific steps for ELM model optimization are as follows: determine the ELM calculation structure and determine the length of the model calculation based on the original number of settlement data; Based on genetic algorithm, the fitness value of each subsidence data individual is calculated to find the optimal solution. Finally, the GA-ELM model is determined and the predicted data is obtained. The specific steps are shown in Figure 2.

Model training and validation
The subsidence of a small water conservancy project is mainly related to the groundwater dynamics, precipitation, air temperature and soil quality changes of the project site.
The dynamic of groundwater is the main factor that affects the settlement of projects. Three activation function of the genetic algorithm to optimize the ELM model, based on four kinds of influence factors of 2 kinds of combinations: 1 enter groundwater dynamic combination, precipitation, temperature and soil changes in four indicators, combination of two input precipitation, temperature and soil changes in three kinds of index, specific combination forms are shown in table 1, we validate the calculation precision of different models under different input methods.

Model accuracy verification
Take root mean square error (RMSE), relative root mean square error (RRMSE), determination coefficient (R 2 ), model efficiency coefficient (E ns ) and average absolute error (MAE) as the model accuracy evaluation index system, and the specific formula is as follows: ∑ | |

Comparison of settlement fitting effect between different models
Comparison and analysis of settlement prediction results and measured values by different models are shown in Figure 3. As can be seen from Figure 3, the simulation accuracy of different prediction models is different, but the overall data change trend is basically consistent with the measured value change trend. The settlement value of different parts of the building is different, the settlement value of the edge of the building is higher than that of the inside of the building. The ELM 1 model performs well in the prediction of building edge and internal settlement, with the simulated value closest to the measured value, while the ELM 12 model has the worst fitting effect.

Comparison of settlement prediction accuracy of different models
The accuracy of the prediction results of different models is shown in Table 2. As can be seen from  Figure 4 shows the GPI index comparison of the prediction results of different models. It can be seen from the figure that GPI is calculated from the simulation results of different models GPI is different at different locations. At the edge of the building, ELM 1 model has the highest GPI (1.904), while ELM 12 model has the lowest accuracy (-3.096). Inside the building, ELM 1 model has the highest accuracy, ELM 12 model has the lowest accuracy, and GPI is 1.052 and -3.948. In summary, GA-ELM sin model is the model with the highest accuracy for settlement prediction of small water conservation projects.

Conclusion
Based on the limit learning machine model, this paper studies the settlement prediction models of small water conservancy projects under different models. Based on the three activation functions of ELM model, genetic algorithm was used to optimize each model respectively. And the calculated value was compared with the measured value. The conclusion was drawn: although there were some differences in the accuracy of different models, the predicted value was basically consistent with the change trend of the measured value. The calculation accuracy of ELM model can be improved by genetic algorithm. Among the 3 activation functions of ELM model, sine activation function has the highest accuracy. In summary, GA-ELM sin model can be the standard model for settlement prediction of small water conservancy projects. In the future research, the model can be compared with the generalized regression neural network model and BP neural network model to further illustrate the scientific nature of the model.