Neural Network Based Predication and Optimization for Groundwater / Surface Water System

Abstract

The present invention relates to a method and apparatus, based on the use of a neural network (NN), for (a) predicting important groundwater / surface water output / state variables, (b) optimizing groundwater / surface water control variables, and / or (c) sensitivity analysis, to identify physical relationships between input and output / state variables used to model the groundwater / surface water system or to analyze the performance parameters of the neural network.

Description

RELATED APPLICATION / CLAIM OF PRIORITY [0001] This application is related to and claims priority from Provisional Application Ser. No. 60 / 347,626, filed Oct. 22, 2001.INTRODUCTION [0002] The present invention relates to a method and apparatus, based on the use of a neural network (NN), for (a) predicting important groundwater / surface water output / state variables, (b) optimizing groundwater / surface water control variables, and / or (c) sensitivity analysis, to identify physical relationships between input and output variables used to model the groundwater / surface water system or to analyze the performance parameters of the neural network. BACKGROUND [0003] Management and protection of water resources around the world is becoming critically important. Overuse and contamination combined with increasing water demand are diminishing the quantity and quality of water available for human use. It has been estimated that by year 2025, over 35% of the world population will face chronic water sho...

AI Technical Summary

Benefits of technology

[0034] In addition, the preferred version(s) of the present invention can provide the following additional benefits:

[0035] (a) The sensitivity analysis of the neural network system provides a better understanding of the natural system and can facilitate the design of a more effective and cost efficient data monitoring and collection system for hydrologic data. The sensitivity analysis results may also be used to improve a numerical model by identifying the interrelations between variables (e.g. precipitation and water levels).

[0039] (e) The neural network system is well suited for using real-time on-line data to initialize the state-transition equations increasing predictive accuracy.

[0041] (g) The neural network derived transition equations can be used to conduct formal optimization in real time using on-line automated data collection systems. The system is able to react adaptively to new data and information and can update optimal solutions on-line.

Problems solved by technology

Overuse and contamination combined with increasing water demand are diminishing the quantity and quality of water available for human use.

Adverse environmental impacts like aquifer overdraft, saltwater intrusion, wetlands dewatering, stream flow depletion, and spreading of groundwater contamination are often consequences of improperly managed systems.

Because of the expenses associated with quantifying these properties, typically a very limited number of field and / or laboratory measurements are made.

Except under relatively ideal hydrogeologic conditions, assigning a “representative” distribution of these parameters across a model grid can be a difficult if not impossible task.

As a result, small-scale features (e.g. lenses or fractures) that can have a significant effect on local or even regional flow regimes are rarely known, much less incorporated into the numerical model.

Even stochastic approaches, which attempt to model the natural heterogeneity of the subsurface conditions, have limited application to real-world conditions because of simplifying mathematical assumptions and relatively sparse field information.

Regardless, there will always be an inherent amount of error between initial conditions assigned to the model and the real-world initial conditions, serving as an additional source of prediction error.

They can be extremely difficult to quantify (e.g. mountain front recharge) and as in the real world they largely determine the pattern of flow in the model.

As with initial conditions, discrepancies between assigned and real-world boundary conditions contribute to prediction error, particularly during transient simulations.

A more fundamental modeling problem may arise if the physical assumptions of the numerical model as represented by its equations do not match the natural system.

Hence model accuracy is questionable.

However, the simplifying assumptions make the problem solvable, which is why they continue to be applied in practice.

Consequently, numerical groundwater flow models are susceptible to producing relatively large predictive errors of water levels and other variables.

This in turns affects the simulation accuracy of predictions made using traditional groundwater / surface water models, and can compromise sound resource management decisions.

There are problems associated with optimization using numerical models, the difficulties of which are discussed below.

A major disadvantage of this method is that the size of the constraint matrix can become extremely large.

However, as it is well known from numerical analysis small step size in numerical differentiation results in high round off error.

However, non-uniqueness of the solutions in physical-based models makes it difficult to evaluate the sensitivity of different factors on the modeled system.

Images