A runoff prediction method, device and electronic device based on hierarchical attention network

The hierarchical attention network method addresses the challenges of predicting runoff by dividing watersheds into sub-watersheds and applying multi-scale attention mechanisms, resulting in improved accuracy and integration of local and global features for precise runoff prediction.

CN118350660BActive Publication Date: 2025-03-18GUIZHOU QIANYUAN POWER CO LTD +2
2 Cites 0 Cited by

Patent Information

Application Number
CN202410429107.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-10
Publication Date
2025-03-18
Estimated Expiration
2044-04-10

AI Technical Summary

Technical Problem

Existing methods for predicting runoff in water resources management are limited by their inability to accurately capture the nonlinear, random, and cyclical characteristics of runoff sequences, which are influenced by natural weather conditions, watershed features, and human activities, leading to inaccurate predictions.

Method used

A hierarchical attention network (HANet) method is employed, which divides a watershed into sub-watersheds, encodes their features, and applies multi-scale attention mechanisms to integrate local and global information, enabling precise runoff prediction by simulating the runoff generation process and accounting for spatial heterogeneity and connectivity.

Benefits of technology

The HANet method enhances prediction accuracy by effectively combining local and global features, overcoming the limitations of existing models in associating local runoff processes with overall spatial connectivity, thereby achieving more accurate runoff predictions in larger watersheds.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present application belongs to the technical field of hydrological prediction, and specifically discloses a runoff prediction method, device and electronic device based on a hierarchical attention network, the method comprising: based on the overall input sequence set of the watershed, dividing the watershed into sub-watershed spaces, and obtaining a first feature sequence set corresponding to each sub-watershed space; converting the first feature sequence set through an encoder to obtain a second feature sequence set, and the second feature sequence sets corresponding to different sub-watershed spaces maintain the same feature number; based on the second feature sequence set corresponding to each sub-watershed space, sequentially performing pooling operations and convolution operations, and merging the input features of the pooling operation and the output features of the convolution operation to obtain sub-watershed space fusion features; based on the overall input sequence set and the sub-watershed space fusion features, performing multi-head attention at the global spatial scale and the local spatial scale, and fusing the multi-head attention representation and the overall input sequence set to obtain a runoff prediction result.
Need to check novelty before this filing date? Find Prior Art