Evaluation of the performance of intelligent neurophasic models and artificial neural networks in predicting and simulating the quality parameter of TDS of rivers (Case study of the Shirin River)

Master's Thesis in Civil Engineering - Hydraulic Structures

Abstract:

Rivers are one of the most important and common sources of drinking, agricultural and industrial water supply. These resources have many qualitative fluctuations due to passing through different platforms and direct connection with the surrounding environment. Therefore, predicting the quality of river flow, which is an uncertain, random and influenced phenomenon of some natural and unnatural factors, plays an important role in the quality management of water resources. According to the defects in the statistical data, the results of the simulation models can be used to discover the defects, correct or complete the data. In order to check the quality status of a water resource, indicators are considered to control the quality of water resources. In order to achieve this, the concentration of dissolved solids (TDS) and electrical conductivity (EC) of Grab hydrometry station located in Ab Shirin river have been evaluated to predict and simulate salinity changes. In forecasting models, monthly delayed inputs of total dissolved solids have been used to estimate salinity by maintaining temporal continuity, and in simulation models, due to the necessity of maintaining temporal continuity and reducing the modeling error, the random combination of total anions and cations has been used as model input. In this study, the intelligent algorithms of artificial neural networks and fuzzy-neural networks have been used to model time series that do not have conditions such as stationarity to apply classical techniques. The results indicate the almost similar performance of the above two methods with acceptable accuracy in modeling the qualitative parameters of the study area. In the end, according to the obtained results, the neurophasic model has less uncertainty in the output values ??compared to the neural network; So that it performs better within the confidence range of most modeling.

Key words: qualitative parameters, prediction, fresh water river, artificial neural networks, simulation, neuro-fuzzy.

Chapter one: basic concepts

1-1 Introduction

One of the most important factors in the development of any region is the availability of quality water sources. Knowing the state of river pollution has made management plans to control river water quality in the future more important. Predicting the quality of river flow in the future, despite being affected by some natural and unnatural factors, plays an important role in managing the quality of water resources.

By predicting the quality of river flow, in addition to managing the exploitation of water resources to meet the needs, and allowing more agricultural and industrial harvests in the time periods when the river is more polluted, it is possible to use diversion routes from the entry of streams with high pollution loads that affect Unfavorable effect on the water quality of the reservoirs. Also, due to the existence of statistical data defects in the quantitative and qualitative data of hydrometric stations, the results of the simulation model of qualitative parameters can be used in order to correct, discover defects, correct or complete the data. Experimental models that try to establish a relationship between input and output data regardless of the parameters used are known as intelligent models. In fact, fuzzy logic, neural computing and genetic algorithms form the foundations of soft computing science. Unlike hard computing [1], soft computing [2] is compatible with the uncertainty in the real world. It is possible to express the basic principles of soft computing in the form of one sentence as follows: "Using the tolerance of inaccuracy, uncertainty and partial truth [3] in order to reach a flexible, solid and low-cost solution" [63]

In the prediction of quality parameters, one can use the time delays of the same parameter, due to its abundance and availability compared to other parameters such as flow rate, temperature, color, etc. used as model inputs. In fact, one of the methods of predicting natural and unnatural processes, including pollution, is to use delayed time series of the same parameter as a predictor. 1- The main goal of this research is to use intelligent neural and fuzzy-neural network models to estimate the salinity of a future time step by examining the influence of monthly delayed time series in the study area.

2- In the following, the problem of simulating TDS using the concentration of different ions in water, pH and flow rate as the input of the models has been investigated and analyzed. The changes of TDS with other qualitative parameters in different rivers have been calculated, among these parameters, total anion and total cation have been selected as inputs to the simulation model, and the results of each model have been discussed and reviewed. is Hydrological forecasts can be divided into short-term and long-term. Short-term forecasts often have a time horizon of several days and are used for warning and real-time operation of water resources systems. In contrast to long-term forecasts, they have a time horizon of more than a week to a year and are used to manage water resources, such as allocating water for irrigation and reducing the effects of drought through water resources management.

Short-term forecasts are usually more accurate and easier to obtain. Mathematical and physical relationships are more important for these predictions and have better simulation capabilities. On the other hand, long-term forecasts have more errors for various reasons and have more complexities in modeling and simulation. Equally, their importance for a water resource management system is very high, so that a small increase in the accuracy of these predictions will bring many benefits to the exploitation system. The first and most obvious benefit of forecasts with long-term time horizons is to make decisions based on water storage and release more dynamic [14]. Therefore, monthly and seasonal forecasts related to river quality parameters and salinity changes are part of long-term forecasts, and the results of these forecasts are very important in managing the quality of water resources.

1-2-1 Modeling for forecasting

The process of modeling for forecasting includes the following steps:

Determining a suitable predictor

Determining a suitable forecasting model

Model calibration

Model validation

1-2-1-1 Determining a suitable predictor

The first step in modeling for Forecasting is the use of a suitable forecaster. The use of appropriate predictors depends on the physical conditions governing the area and the study area.

Indicative variables that are used to predict the stream quality include:

The flow rate of the stream in the past periods of time, the electrical conductivity of EC and the total dissolved solids TDS, and it also includes the rest of the measured quality parameters of the stream.

The general form of the equations obtained based on these variables is It is as follows:

Q = f (X1 , X2 , X3 , … , Xn )

where Xi is the i-th index variable among n variables and Q is the stream quality or stream salinity parameter in the desired time period of the forecast.

Three methods of modeling hydrological phenomena[19]

Inspection and analysis of various phenomena in the field of water and environmental engineering, such as quality management of water resources, according to the requirements of the plan (importance, desired accuracy, facilities and time) can be done in the form of the following three general methods:

Numerical methods

Experimental methods (Laboratory and field)

Example-based learning methods (artificial intelligence or data mining)

In recent decades, soft computing tools and intelligent systems have been introduced as new methods of modeling complex engineering systems. The basis of this method is summarized in the two categories of statistics and artificial intelligence, which artificial intelligence methods are considered as machine learning methods. These methods actually determine the relationship between dependent and independent parameters and somehow fit the most suitable function on them and are able to approximate any non-linear function [11], [47].