CN-121981421-A - Method and device for predicting water demand of non-metering residential area based on opposite domain adaptive neural network

CN121981421ACN 121981421 ACN121981421 ACN 121981421ACN-121981421-A

Abstract

The invention discloses a method and a device for predicting water demand of an unmetered residential community based on a reactive domain adaptive neural network, wherein the method comprises the following steps of 1) obtaining water consumption data of a source domain and a target domain, wherein the water consumption data comprises building environment and social grading indexes (BESS indexes) of the residential community and corresponding water consumption data, 2) constructing an initial model comprising three parts of an encoder, a decoder and a domain countermeasure, 3) training the initial model by adopting the water consumption data, and carrying out 0-value mask on water consumption mode characteristics of the metered residential community in the target domain to simulate an unmetered scene during training to obtain a prediction model. The invention also provides a device for predicting the water demand of the non-metering residential area. The invention can realize the accurate estimation of the water demand of the non-metering residential area by only relying on the BESS index of the residential area, and solves the prediction problem caused by water data loss in a non-metering scene.

Inventors

MA XINZHEN
Zou Zekun
LONG ZHIHONG
XU GANG
ZOU KANGBING
YU TINGCHAO
SHAO YU

Assignees

广州市自来水有限公司
浙江大学
浙江大学长三角智慧绿洲创新中心

Dates

Publication Date: 20260505
Application Date: 20251205

Claims (8)

1. The method for predicting the water demand of the non-metering residential area based on the reactive domain adaptive neural network is characterized by comprising the following steps of: Step 1, acquiring water consumption data of a source domain and a target domain, wherein the water consumption data comprises BESS indexes of a metering residential district in the source domain and the target domain, water consumption data corresponding to the BESS indexes and BESS indexes of a non-metering residential district in the target domain; Step 2, constructing an initial prediction model based on a reactive domain adaptive neural network, wherein the initial prediction model comprises an encoder, a decoder and a domain countermeasure; the encoder integrates a BESS feature extraction module, a water use mode extraction module, a self-attention migration module and a time domain alignment module; the BESS feature extraction module is used for extracting data features in the BESS feature indexes of the residential quarter so as to obtain corresponding BESS features; The water consumption pattern extraction module is used for projecting water consumption data to a Fourier space for feature extraction, generating corresponding water consumption pattern features, and fusing BESS features and the water consumption pattern features to obtain fused features; The self-attention migration module processes the fusion characteristics based on a self-attention mechanism, takes the BESS characteristics as migration media, realizes migration of the water pattern characteristics from the metering residential area to the non-metering residential area, and outputs migration characteristics; the time domain alignment module is used for executing convolution operation on the time dimension of the migration characteristic, eliminating the dislocation deviation of the water consumption mode of the source domain and the target domain on the time dimension, and obtaining a time domain invariant characteristic; The decoder adopts a residual error connection network structure, decodes the time domain invariant feature into water consumption variable quantity, and outputs water consumption estimated values of the metering residential area and the non-metering residential area; The domain countermeasure adopts a residual error connection network structure to carry out domain classification on the time domain invariant features of different residential areas and output the domain classification result of each residential area; And step 3, training the initial prediction model by adopting the water consumption data, and carrying out 0-value mask processing on the water consumption mode characteristics of the metering residential area in the target domain in the training process so as to simulate the characteristic missing scene of the non-metering residential area, thereby obtaining the water consumption prediction model of the non-metering residential area after the training is completed.
2. The method for predicting water demand of non-metered residential cells based on a reactive domain adaptive neural network according to claim 1, wherein the BESS indexes comprise three dimensional climate indexes of time of day and monday, air temperature, humidity and weather conditions, three dimensional residential cell social attribute indexes of water bill, construction time and room price, and four dimensional residential cell spatial attribute indexes of cell occupation area, number of buildings in a cell, average building occupation area and building elevation.
3. The method for predicting water demand of non-metered residential areas based on a reactive domain adaptive neural network according to claim 1, wherein the BESS feature extraction module extracts complex nonlinear interactions between BESS indexes with a residual connection network.
4. The method for predicting water demand of non-metered residential communities based on the reactive domain adaptive neural network according to claim 1, wherein the water pattern extraction module maps water data to a frequency domain through fourier transformation to generate corresponding sine eigenvectors and cosine eigenvectors; And splicing the sine characteristic vector and the cosine characteristic vector, and performing deep treatment through a residual error connection network to obtain corresponding water use mode characteristics.
5. The method for predicting water demand of non-metered residential communities based on the reactive domain adaptive neural network according to claim 1, wherein the self-attention migration module performs axis exchange operation on the input fusion characteristics and performs three different linear changes based on the fusion characteristics after the axis exchange operation so as to obtain corresponding query tensors, key tensors and value tensors; Based on the query tensor, performing attention weight calculation on the key tensor and the value tensor to obtain corresponding similarity weights, and normalizing the calculation result into probability distribution by using a softMax function; the value tensor and the similarity weight are multiplied to output a similarity-based migration feature.
6. The method for predicting water demand in non-metered residential cells based on a reactive domain adaptive neural network of claim 1, wherein the time domain alignment module employs one-dimensional convolution to enhance translational invariance of migration features in a time dimension.
7. The method for predicting water demand of non-metered residential areas based on a reactive domain adaptive neural network according to claim 1, wherein in the water pattern extraction module, For an effective water pattern feature extracted from water usage data of a metered residential area, and for a non-metered residential area, the water pattern feature is filled with 0 values.
8. A non-metered residential-area water demand prediction apparatus, characterized by the steps for executing the non-metered residential-area water demand prediction method based on the reactive domain adaptive neural network as claimed in any one of claims 1 to 7.

Description

Method and device for predicting water demand of non-metering residential area based on opposite domain adaptive neural network Technical Field The invention belongs to the technical field of municipal engineering and urban water supply networks, and particularly relates to a method and a device for predicting water demand of an un-metered residential community based on a reactive domain adaptive neural network. Background The accurate estimation of the water consumption of urban users is crucial to realizing efficient allocation management of urban water resources, can provide data support for water distribution scheme optimization, water supply network leakage and pollution monitoring, and lays a foundation for water resource long-term operation strategy formulation. However, due to the limitation of technology and economic conditions, most urban users are not provided with water consumption metering equipment, so that the urban water distribution process lacks a real-time accurate sensing means, and the fine development of water resource management is restricted. In order to reduce uncertainty of water resource allocation, researchers have developed a Water Distribution System (WDS) hydraulic state inference method in the prior art to obtain core parameters such as node water demand, pipeline flow or node pressure. Traditional hydraulic state inference methods often combine hydraulic simulation models (e.g., EPANET) with sensor monitoring data to extrapolate non-metered user water demand back by minimizing deviation of the simulation values from the measured values. Such methods are mostly based on bayesian framework construction, typically such as kalman filter recursive estimation, variational bayesian inference to maximize posterior probability, etc. The patent document CN120450158A discloses a water supply network full-flow water demand prediction method based on a space-time diagram neural network, and the method can construct a full-connection adjacency matrix of sensor nodes with corresponding scales based on time steps to be detected and input data in the training process by utilizing a self-adaptive diagram learning module. The patent document CN116822244A discloses a water supply network water demand checking method based on cut-off normal distribution (belongs to the field of pipe network hydraulic modeling), which comprises the core steps of (1) adopting the cut-off normal distribution to construct cut-off prior probability distribution of node water demand and cut-off likelihood function of monitoring data, (2) fusing the distribution and the function based on Bayesian theorem to construct cut-off posterior probability distribution, establishing a checking objective function by maximizing the distribution, (3) adopting a Newton iteration method to solve the objective function, obtaining node water demand adjustment quantity and iterating to obtain final water demand. The method improves the checking precision by limiting the water demand distribution interval, but has the obvious defects that firstly, the time independence assumption based on Markov property is easy to lose systematic characteristics of a water consumption time mode, and secondly, the method depends on single numerical values such as historical water cost and the like as prior constraints, so that the method is difficult to adapt to complex water consumption behaviors in a dynamic urban environment. Disclosure of Invention The invention aims to provide a method and a device for predicting water consumption of an un-metered residential area based on a reactive domain adaptive neural network, and the method can accurately estimate the water consumption of the residential area lacking metering equipment only by relying on the constructed building environment and social attribute characteristic indexes of the urban residential area. In order to achieve the first purpose of the invention, the invention provides a method for predicting the water demand of an un-metered residential area based on a reactive domain adaptive neural network, which comprises the following steps: Step 1, acquiring water consumption data of a source domain and a target domain, wherein the water consumption data comprises BESS indexes of a metering residential district in the source domain and the target domain, water consumption data corresponding to the BESS indexes and BESS indexes of a non-metering residential district in the target domain; Step 2, constructing an initial prediction model based on a reactive domain adaptive neural network, wherein the initial prediction model comprises an encoder, a decoder and a domain countermeasure; the encoder integrates a BESS feature extraction module, a water use mode extraction module, a self-attention migration module and a time domain alignment module; the BESS feature extraction module is used for extracting data features in the BESS feature indexes of the residential quarter so as to obtain corresponding BE