CN-122021253-A - Short-term prediction method for urban water supply capacity driven by KAN-PINN under multi-factor coupling

CN122021253ACN 122021253 ACN122021253 ACN 122021253ACN-122021253-A

Abstract

The invention relates to the technical field of intelligent water affairs and artificial intelligence and discloses a short-term prediction method of urban water supply driven by KAN-PINN under multi-factor coupling, which comprises the steps of acquiring and preprocessing historical water supply and multi-factor data, and screening key factor information to construct a solution training test set; and (3) constructing a KAN-PINN cooperative model, extracting nonlinear characteristics by a KAN network, introducing physical constraints reflecting monotonicity relations into a loss function, optimizing the loss function by using a training set to complete model training, and inputting a test set into the model to obtain a short-term water supply prediction result. The method solves the problems of insufficient multi-factor strong coupling modeling and insufficient fusion of a data mechanism, and improves the short-term water supply prediction precision and model robustness.

Inventors

WANG TINGTING
LU XUEHUI
WANG HONGZHI
LIU WENZHENG

Assignees

新疆河润科技股份有限公司

Dates

Publication Date: 20260512
Application Date: 20251226

Claims (10)

1. A short-term KAN-PINN co-driven urban water supply prediction method under multi-factor coupling, comprising: Acquiring a historical data sequence containing historical water supply and a plurality of influencing factors, and preprocessing the historical data sequence to obtain a regular data sequence; screening key influence factors associated with water supply quantity prediction from the regular data sequence, and constructing a training set and a testing set containing the key influence factors; Constructing a KAN-PINN collaborative prediction model, extracting nonlinear characteristics of the key influence factors through a KAN network structure of the KAN-PINN collaborative prediction model, and introducing physical constraints in a loss function of model training, wherein the physical constraints reflect a monotonicity relation between the key influence factors and water supply; Training the KAN-PINN collaborative prediction model by using the training set, and updating model parameters by optimizing a loss function containing the physical constraint; And inputting the test set into the trained KAN-PINN collaborative prediction model to obtain a short-term water supply quantity prediction result.
2. The multi-factor coupled KAN-PINN co-driven short-term city water supply prediction method of claim 1, wherein the step of preprocessing the historical data sequence comprises: Setting the length and the step length of a sliding window, calculating the mean value and the standard deviation of data in the window, marking data points exceeding the range of three times of the mean value and the standard deviation as candidate abnormal values, and carrying out secondary verification and final elimination on the candidate abnormal values based on the physical boundary constraint of a water supply system and the known scene record.
3. The method for short-term prediction of KAN-PINN co-driven urban water supply under multi-factor coupling according to claim 2, further comprising the steps of preprocessing the historical data sequence, dividing the preprocessed effective data into a training set and a testing set by adopting a random forest interpolation method, training a random forest model by utilizing the training set, optimizing model parameters by grid search, and performing prediction interpolation on the missing values by utilizing the optimized model.
4. The method for short-term predicting urban water supply driven by KAN-PINN under multi-factor coupling according to claim 1, wherein the process of screening the key influencing factors related to water supply prediction includes quantifying linear correlation strength of each influencing factor and water supply by Pearson correlation analysis, and performing preliminary screening according to a preset correlation coefficient threshold to obtain a candidate factor set.
5. The method for short-term prediction of urban water supply driven by KAN-PINN under multi-factor coupling according to claim 4, wherein the process of screening the key influencing factors further comprises the steps of analyzing factors in the candidate factor set by two by adopting a coupling co-scheduling model, calculating coupling degree and co-scheduling among the factors, constructing a coupling strength matrix, and identifying a high coupling factor group based on the coupling strength matrix so as to determine the key influencing factors.
6. The short-term prediction method of urban water supply driven by KAN-PINN under multi-factor coupling according to claim 1, wherein in the KAN-PINN collaborative prediction model, a KAN network structure for extracting nonlinear characteristics is a multi-layer network including an input layer, at least one hidden layer and an output layer, and an activation function adopted by the hidden layer and the output layer is a learnable B-spline basis function.
7. The multi-factor coupled KAN-PINN co-driven urban water supply short-term prediction method of claim 1, wherein the model trained loss function is expressed as a total loss function, the total loss function being a weighted sum of a data driven loss term, a structural loss term and a physical inconsistency loss term, wherein the structural loss term is used to constrain network weights to prevent overfitting, and the physical inconsistency loss term is used to quantify the degree of deviation of model predictions from the monotonicity relationship.
8. The short-term prediction method of the urban water supply capacity driven cooperatively by KAN-PINN under the multi-factor coupling according to claim 7, wherein the physical inconsistency loss term is calculated based on artificially generated sample data, and for each of the key influencing factors, the calculation model outputs a partial derivative sign of the factor, and compares the partial derivative sign with a preset monotonicity factor reflecting the physical rule, and accumulates penalty terms of inconsistent signs.
9. The short-term prediction method of the urban water supply capacity driven cooperatively by KAN-PINN under the multi-factor coupling according to claim 1, wherein after the short-term water supply capacity prediction result is obtained, the prediction performance of the KAN-PINN cooperative prediction model is evaluated by means of a mean square error, a root mean square error, an average absolute error and a fitting degree.
10. The short-term prediction method of the urban water supply capacity driven by the KAN-PINN under the multi-factor coupling according to claim 1, wherein the method further comprises a generalization capability verification step of acquiring water supply capacity of another city or region and related influence factor data as a new test set, predicting the new test set by using the KAN-PINN collaborative prediction model completed based on the original training set, and calculating a prediction performance index to verify the trans-regional applicability of the model.

Description

Short-term prediction method for urban water supply capacity driven by KAN-PINN under multi-factor coupling Technical Field The invention relates to the technical field of intelligent water affairs and artificial intelligence intersection, in particular to a short-term prediction method for urban water supply capacity driven by KAN-PINN under multi-factor coupling. Background With the deep advancement of the urban process and the aggravation of climate change, the urban water supply system is used as a core infrastructure for maintaining the economic and social operation of the city, and the dispatching optimization and the efficient allocation of water resources are directly related to the life quality guarantee of residents, the stable operation of industrial production and the sustainable development of ecological environment. The short-term water supply quantity prediction is used as a key technical link of intelligent water service construction, can provide decision-making basis for accurate regulation and control of water supply pump stations, dynamic balance of pipe network pressure and scientific formulation of emergency water supply plans, and is an important premise for realizing safe, economical and efficient operation of a water supply system. However, under the current urban development background, short-term water supply is affected by meteorological condition dynamic change, user water behavior mode transition, holiday effect periodic fluctuation, industrial structure adjustment and other multi-factor interweaving, obvious nonlinearity, time variability and uncertainty characteristics are presented, and higher requirements are put on the accuracy and robustness of a prediction model. Traditional prediction methods are mainly based on statistical models or single machine learning algorithms. Although the autoregressive integral moving average (ARIMA) time series analysis method can describe the linear trend of data, the complex nonlinear relation under the multi-factor coupling effect is difficult to capture, and the prediction deviation is obviously increased when the working condition is abrupt. In recent years, deep learning models such as a cyclic neural network (RNN) and a variant long-short-term memory network (LSTM) thereof, a gate-controlled cyclic unit (GRU) and the like are widely applied in the field of water supply prediction by virtue of strong sequence modeling capability. In the prior art, the method has the advantages of researching the long-term and short-term dependence of a water supply time sequence captured by an LSTM model to realize the high-precision prediction of ultra-short time water demand, researching the dynamic change mode of learning the water supply sequence by utilizing a self-Attention mechanism based on a transducer architecture, researching and fusing LSTM, GRU and the Attention mechanism to construct a mixed model, or providing a CNN-BiLSTM-Attention superposition model to enhance the local feature extraction and time sequence modeling capability. However, the data driving model has three general limitations that firstly, the model is excessively dependent on the quality and the scale of a training sample, is easily interfered by abnormal values and lacks internal constraint on physical laws such as water conservation, supply and demand balance and the like, so that the model has insufficient interpretation, secondly, the coupling strength among multiple factors is not effectively quantized, hidden relations among variables such as weather, society and economy are difficult to accurately describe, thirdly, the generalization capability is obviously reduced under the conditions of data scarcity or extreme working conditions, and the prediction stability is difficult to guarantee. The physical information neural network (Physics-Informed Neural Networks, PINN) realizes the organic fusion of the data driving and mechanism models by taking a physical equation as a constraint embedding loss function, is successfully applied to the fields of fluid mechanics, energy scheduling and the like, and effectively improves the interpretation and extrapolation capability of the models. However, in the water supply quantity prediction scene, a single PINN is faced with the problem of insufficient structural flexibility, the fixed network architecture has limited fitting capability on high-dimensional complex nonlinear characteristics generated by multi-factor strong coupling, and deep interactive relations among multi-source heterogeneous data such as weather, user behaviors and the like are difficult to fully mine. As an emerging neural network structure, kolmogorov-Arnold Networks (KAN) are built based on Kolmogorov representation theorem, and a learnable activation function is adopted to replace a traditional fixed activation function, so that the Kolmogorov-Arnold network has better precision and parameter efficiency in a high-dimensional function approximation task. T