KR100540009B1 - Water need forecasting method using neural network - Google Patents

Abstract

The present invention relates to a water demand prediction method using a neural network. The present invention utilizes neural networks and evolutionary algorithms to secure a stable supply by reasonably connecting and operating water intake, water purification and pressurization plants by reflecting typical changes due to demand fluctuations, weather conditions, seasonal factors, etc. Constructing an intelligent model of water demand forecasting; constructing demand forecasting software on the upper computer using the neural network model-based predictive control method based on the intelligent model; Demand forecasting software uses the backpropagation learning algorithm based on the operational data collected from the system and the data from the meteorological database. Building stages, Using a neural network model based predictive control algorithm, from a model by a step of determining the predicted demand in the current demand and the weather conditions. According to the present invention, it is possible to secure a stable supply by reasonably linked operation of the intake, water purification and pressurization plants by reflecting the usual changes due to water demand fluctuations, weather conditions, seasonal factors, and the like.

Description

Water demand prediction method using neural network {WATER NEED FORECASTING METHOD USING NEURAL NETWORK}

The present invention relates to a water demand prediction method using a neural network.

In general, water supply facilities are used by all people living in urban life and are important facilities to maintain daily activities and daily activities of the city.The lifeline corresponds to the population concentration of the city, improvement of living standards, and economic growth. line), it is required to stably supply safe and high quality water.

To maintain this, the water management operation (hereinafter referred to as "water operation"), which is a comprehensive maintenance plan for waterworks based on drainage control, remains an important task.

The system supporting water operation is to control and monitor all the water facilities from the water purification plant to the consumer to efficiently supply safe water in terms of quantity, water pressure and water quality. This forecast makes it possible to establish an operational plan for water supply to facilitate the operation of rational water pumps, the adjustment of valves, and the setting of reservoir levels.

Therefore, for optimal operation of drainage facilities, it is necessary to understand the demand structure of water demand considering the characteristics of the target area and to accurately predict demand.

Domestic water management technology (hereinafter referred to as "water management system") has been developed, installed and operated in earnest from the 1990s, based on the systems of developed countries such as Japan, with reference to large corporations. The steps to utilize the results are as follows. ① Calculate daily demand by system using various demand forecasting techniques. ② Calculate forecasted demand by time by analyzing demand patterns using relative time coefficients (aka: demand pattern simulation of data granulization method). ③ The calculated demand forecast by hour calculates the amount of water delivered by branch point to provide economic operation, and provides the operation plan information. ④ Prepare the pump switching table for each hour by considering the electricity charges in the forecasted demand and provide it to the operator. .

Generally, the data granulization technique used to calculate the forecasted demand by time is generally used, but the multiple regression, ARIMA, and Kalman filter model techniques are used as the models used to estimate the daily demand.

However, there has been much debate on which technique yields the most reliable demand forecasting results, and water demand forecasting has been devised for the purpose of controlling drainage. Water demand that can provide important information There are not many system softwares for short-term forecasting, and among them, no neural network model that uses water structure as a black box has been applied. The development of is urgently needed.

Accordingly, the present invention provides a water demand prediction method using a neural network to determine the optimal forecast demand by using intelligent modeling and nonlinear model predictive control technique for stable operation of the water management system, the forecast demand and the actual demand of the intelligent model The present invention provides a method for achieving the output prediction precision goal of the intelligent model, which sets the average% error of the error to less than or equal to 10%.

In addition, the past operation data of the target branch for manual operation, which is indirect information, is selected as a standard for demand forecasting and stabilization target, so that the operation can be performed by online setting of optimal demand forecasting. An object of the present invention is to provide a water demand prediction method using neural networks.

Water demand prediction method using a neural network according to the present invention for achieving the above object is in the conventional water demand prediction method, the water intake system reflects the usual changes due to changes in demand, weather conditions, seasonal factors, etc. In order to secure a stable supply by rationally linking the water purification plant and the pressurization plant, constructing an intelligent model of water demand prediction using a neural network and an evolutionary algorithm, and using a neural network model-based predictive control technique based on the intelligent model. Configuring demand forecasting software to calculate the demand for the target water by using the host computer, connecting the host computer of the operating system, a database containing weather information, and the like through a LAN network, and the demand forecasting software system Operational data collected from, weather DB Constructing a demand prediction neural network model using a backpropagation learning algorithm based on the data, and determining the current demand and forecast demand in weather conditions using a neural network model-based predictive control algorithm from the intelligent model. Characterized in that made.

In addition, the optimal forecasting demand is determined by constructing an optimal demand forecasting model using intelligent modeling techniques such as neural networks and evolutionary algorithms and the nonlinear model predictive control technique based on the intelligent process model. It is another feature that the process simulator is configured in a host computer.

Hereinafter, exemplary embodiments of the present invention will be described with reference to the accompanying drawings.

Water operation refers to the water supply facilities standard ('97 .12) Chapter 9 4.1 General provisions, "Planting the application of measurement and control in water supply, plant control for individual facilities such as water intake, water purification plant and drainage, and drainage from water sources It is the same as the system operation control in "There is a system operation control that integrates and manages the entire water supply facility to the facility." In other words, it does not mean that each plant or facility is operated from intake to water purification and drainage separately, but the integrated operation and management of all facilities of waterworks.

In the water operation system for realizing such water operation, the operation data is generated and various policies are generated to satisfy the three main goals of water supply, water quality, quantity, and pressure. In order to respond to changes in demand, operate facilities scientifically and economically, and ensure equal water supply in each water supply area by early detection of water supply accidents, rapid recovery, and linkage of water intakes. The water quality within drinking water quality standards is secured and maintained by comprehensively managing individual water quality such as water purification plant, pressurized plant, and drainage basin. In other words, with the scientific and reasonable operation, the ultimate goal is to improve the quality of work for the entire water supply business, such as demand forecasting and stable water supply through equalized water.

1 shows an overview of the key operating system functions. In order to establish an efficient water supply control system, it is important to forecast the water consumption of consumers and to predict the water consumption of the consumer, and to plan stable supply and control the supply based on this. In other words, it is the basis of economic operation that determines the amount of supply by utilizing the facility capacity of the drainage or purified water without any inconvenience in using the water of the consumer, and controls to supply only the determined flow rate. Therefore, it is necessary to adjust the branch valve so that the flow rate to each customer (the branch point) is supplied in consideration of the operation status of the recipient facility (drainage level or water level). When the supply flow rate of customer (branch point) is determined, the inflow and outflow volume is determined in consideration of the characteristics of the facility (water treatment plant, water intake plant, pressurization plant), power rate, and operating conditions. Considering the characteristics, the optimum pump operation number is determined.

Water Demand Forecasting Daily (short-term forecast) water demand forecasting by water supply zone is performed, and the hourly demand forecasting by drainage pattern analysis is performed as a basic function. However, long-term forecasting is performed on a monthly basis, and the calculated hourly demand forecast is based on the electricity bill, drainage level, water level, production volume, flow rate for each water supply zone, and smoothing inflow to ensure economic operation. Calculate the two-day operation plan, transfer it to the host computer, and the reservoir operation plan that compensates for the case where the water level is out of operating range or the water level is out of tolerance due to demand forecast error and accident. The function must be performed. In addition, external data (weather, climate, etc.) for demand forecasting should be entered automatically.

Appropriate techniques must be applied to build a demand forecast model, which typically includes: 1) historical performance drains, 2) the next day's highest expected daytime temperatures, 3) weather, and 4) days of the week (weekdays / holidays / special days). Calculate the demand forecast amount by considering Drainage pattern analysis uses appropriate techniques that take into account 1) seasons, 2) days of the week (weekdays / holidays / special days), and 3) weather items.

The operation flow for the main functions of daily demand forecasting is as shown in Figure 2, and the function predicts 1-2 days of daily demand by water supply area, and establishes the operation plan and establishment of water supply plan for pressurized water, water purification plant, and intake station. It is a function that calculates the quantity demanded. Historical performance volume collected from the host computer's database by daily time triggering or manual start of the driver, weather information for the next day (climate, expected peak temperature during the day, etc.) and date information (xxx, xx, xx, xx, xx) Collect it to predict the next day's drainage. Prediction uses algorithms such as computational models such as regression model, Kalman filter method and fuzzy prediction model, which are adaptive techniques that change the parameters of the prediction model according to the prediction error. Corresponds.

Although there are many methods for estimating daily demand in daily water supply systems, the multiple regression model, transition function ARIMA, and Kalman filter are generally used. The multiple regression model is the most common linear model that formulates the direct effects of external factors (such as past multiple days of performance, the next day's peak temperature, weather, and days of the week (weekdays / holidays / special days)). It is a linear model that considers the effects of indirect effects and time delays, as well as the direct effects of external factors (eg, performance multiples of the past few days, the highest expected daytime of the next day, weather, and days of the week (weekdays / holidays / special days)). The Kalman filter model is a function that expresses non-linear relations of external factors such as past multiple days of performance, the next day's peak temperature, weather, and day of week (weekdays / holidays / special days).

The multiple regression model develops a daily water demand forecasting model using the factors considered to affect daily water demand as independent variables, and water demand as a dependent variable. In the case of a linearly significant effect, the linear combination of the output sequence and multiple input sequences is called a regression model. Here, the output series is called the objective variable (dependent variable), and the input series is called the explanatory variable (independent variable). This is expressed as equation (1).

The multiple regression model undergoes three stages of verification to establish the model in a general way that is applied to a wide variety of fields. The first is about the independent variables, and it is verified whether they are valid as independent variables and whether they are statistically independent from each other. Whether it is valid as an independent variable is verified through a t-test of 95% significance level, and whether it is independent between each other is verified through a variance expansion factor. The second is about whether the residual generated by the model is an error that occurs randomly, and is verified through the Dublin Watson statistic that indicates the autocorrelation of the residual. Thirdly, the predicted value of the model is correct and verified by correlation coefficient and absolute mean error.

Transition Function The ARIMA model is a function in which the output sequence is linearly affected by the input sequence in a dynamic system. It consists of two parts. One is the transition function model. The influence of the input sequence is modeled through cross-correlation analysis and pre-whitening. The other is a term representing the effects other than the input sequence and the delay of the input / output sequence. It is modeled with ARIMA. After the cross-correlation analysis, the process leads to the order estimation and parameter estimation of the model, and the basic equation is shown in Equation (2).

The transition function model and the ARIMA model can be expressed as Equation (3).

Research using transition function ARIMA model in the field of environment is diverse, such as traffic noise and daily water demand prediction. The transition function ARIMA model statistically verifies the significance of the model by selecting each parameter, determining the order, and estimating the parameters.

The index that selects the most optimal model is verified by Chi-square test, AIC and SBC as follows. The chi-square test verifies the autocorrelation of the model residuals, and the AIC and SBC make the most accurate predictions with the minimum parameters.

The chi-square test value determines whether the residual follows the white noise assumed in the model, and is determined as the Portmanto test value of Equation (4).

The Portmanto test above is approximate to the chi-square test of degrees of freedom (Kpq; p: AR course order, q: MA course order), and the Portmanto test Q above is 95% of the chi-square test. A judgment is made that meets the significance level. Therefore, it is significant if the value of the chi-square statistic is 0.05 or more.

The values of Akaike information criterion (AIC) and Schwartz's Bayesian criterion (SBC) are shown in Equations (5) and (6), respectively.

AIC = -2 × maximum log likelihood + 2 × number of fit parameters

Where n is the effective length of the time series

AIC and SBC are in two parts. One is the maximum, indicating the accuracy of the model and having a negative value. The other is to have a positive value as the number of fit parameters. This is a compromise between the model so that it has an accurate predictive value with a minimum fit parameter. The smaller the value, the better the model.

In general, the system is subject to control inputs and disturbances, and the state of the system is observed through the observation device. Observations are often scattered by noise and are not directly observed. In this situation, the Kalman filter model extracts information about the state of the system from the observed data. This method assumes the statistical properties of noise, taking note of the dynamic characteristics of the system generating the time series output, and sequentially estimates the state of the system using the initial value information and the data observed at each time. same.

This Kalman filter is a method of optimally estimating the amount of state at t based on observation data of input and output variables at time t in Discrete Linear Dynamical System. It is described by a differential equation such as

Here, Equation (7) is called a state equation, and Equation (8) is called an observation equation, and is expressed as shown in FIG. Where X _k represents the system state at time tk and is called the state vector. Z ^-1 of the system represents the step delay operator and converts X _{k + 1} one point forward (X _k = z ^-1 X _{k + 1} ). The process of the system is assumed to be Gaussian. The state X _{k + 1} before the point in time is represented by a dynamic linear system that takes as input the system matrix B _k at the point _k and the captured W _k of the system.

In the Kalman filter model of the above matrix representing the changes in the state value X _k B _k and the observation matrix representing the input-output relationship of the system A _k is set to the physical structure of the application of the model to the default, the observations X _k to FIG. The state quantity Q _k is estimated through a loop of four.

Like the human brain, the neural network described in the present invention efficiently processes complex data and has a learning ability on the relationship between input and output. The neural network is composed of several neurons and connection loads connecting them, and the basic structure of the neural network shown in FIG. 5 is connected to several input channels and one output channel to one neuron. A single neuron accepts the output of the different neurons from each input channel and the product of each connection load as inputs, and processes the sum of these inputs using a sigmoidal activation function and generates one output. If the output channel of the neuron i and the input channel of the neuron j is connected, there is a connection load Wij between the connected channel, the neural network may have a multi-layer structure connected by the connection load.

The characteristics of a multilayer neural network including an input layer, an output layer and one or more intermediate layers are as follows.

end. I / O data can be processed in parallel to increase computational power.

I. As it has adaptive learning ability by associative reproduction and learning, it can effectively deal with problems that mathematical algorithms cannot apply.

All. The addition of I / O data enables continuous performance improvement.

la. When applied by the control method, it can be processed in real time by massively parallel processing.

hemp. It has the robustness against the change of the control environment or the noise by the adaptive learning ability.

bar. No environment or model is required and no explicit control rules are required.

four. Damage to some of the neurons is less likely to affect the system's performance.

Among the various neural network models, the model used for modeling water demand prediction is the BP (Back-propagation) neural network model. The BP neural network proposed by Rumelhart, Hinton et al. Has a structure as shown in FIG. 6 and determines the connection load through repetitive learning. The BP learning algorithm is an algorithm that learns the connection load of the forward link used in the pattern learning process.

In FIG. 6, the lowest layer is an input layer, the highest layer is an output layer, and is composed of a hidden layer between them. All neurons take as input the product of the neurons of the lower layer than their own layer and the product of each linking load, and use the sum of these inputs to calculate the output using the sigmoid function and send this output to the higher layer of neurons. It has a structure to convey. The learning process is performed with a set of input and output data obtained in the process, and in the structure of the neural network, as shown in FIG. 6, the input / output layer and the middle layer have input terminals and output terminals, respectively.

The input of the input layer accepts the input data obtained through experiments, and the output of the input layer is a normalized value of the input data. The input of the middle layer is the sum of the product of the connection load and the normalized input, as shown in Equation (12).

In addition, the output terminal of the middle layer is represented by Equation (13) as a sigmoid function value of Equation (12).

The input and output stages of the output layer are obtained by equations (14) and (15) using the values obtained at the output stages of the intermediate layer and the connection loads.

If the output of Eq. (15) calculated in the upper direction and the output pattern obtained in the experiment match, the learning is terminated. If not, the connection load value is corrected using the gradient descent method in the lower direction to reduce this difference. Correction of connection load Made by , Are represented by equations (16) and (17). Here, the learning rate has a value of 0 <η <1.

Equation (18) is the difference between the output inferred by the iterative learning and the actual output data. If this value is within a certain range, the learning ends and the modeling ends.

Among the BP neural network models, an algorithm used to construct a process model in the present invention is an error back-propagation algorithm. The error backpropagation algorithm can learn from the input and output data representing the characteristics of the target nonlinear function to be learned as a basic pattern to mimic the functional characteristics represented by this pattern, and as a result of this learning, other similar general For input, it has a generalization that reacts similarly to the output of a real target function. The multi-layer neural network structure for the error backpropagation algorithm usually used for the control uses one hidden layer.

Therefore, the basic structure of the neural network to be used in the present invention is composed of an input layer, an intermediate layer, and an output layer. At this time, each layer is composed of nodes corresponding to neurons of the neural tissue, and each layer is connected to each other by a connection weight. The node itself has an actuation function or a logistic function. This study uses a sigmoid type function such as Equation (8).

This function has a value between 0 (FALSE), 1 (TRUE), or 0 and 1, depending on the range of incoming inputs, which allows you to connect neurons between layers, logically delimiting the ambiguity of the input pattern. Note changes the load vector to induce the neural network learning and adaptation.

At this time, the change direction of the load vector is the direction of minimizing the difference between the output O (k) and the target value T (k) of the neural network in the current state as the cost function J or the performance index. .

At this time, the change formula of the load is

Where η is the learning rate and is a parameter that determines the rate at which the connection load changes.

In addition, we use a modified algorithm to improve the learning time and to avoid local minimum values. In the present invention, we add the norm of the rate of change of the error to the existing method of constructing the cost function as the norm of the output error. Add performance to improve overall algorithm performance.

In this case, the rate of change of the error can be expressed as

The adjustment of the neural network load vector is modified as follows.

η1 is the learning rate for the output error and η2 is the learning rate for the change in error.

This method has the advantage of increasing the learning speed, avoiding the local minima at high energy level and converging to the minimum value at low energy level compared to the conventional error backpropagation algorithm.

In optimizing the neural network model, factors affecting the convergence performance of the neural network include the structure of the neural network (number of middle layers, number of neurons in the middle layer, etc.), learning parameters (learning rate, momentum coefficient, range of initial connection load, Learning tolerances, the slope of the sigmoid function, etc.). Since the optimal values of these elements are not known at the beginning, it is difficult to find the optimal neural network.

Genetic algorithm, a kind of parallel search, is used to optimize the neural network model by searching for the optimal values that affect the convergence performance of neural networks. Genetic Algorithm is a search algorithm based on natural selection and genetic properties. It encodes the parameters to be optimized into finite length strings and finds the optimal solution through the process of selection, crossover, and mutation for generations. The application of the genetic algorithm to search for the optimal neural network is performed through the following steps.

As a first step, randomly generating a group of character strings that are candidates for the optimal solution at generation of the initial generation. Here, the string is formed by coding in binary as shown in FIG. 7 and the number of individuals in the group is set in consideration of the tolerance.

The second step is to calculate the fitness of each individual, and to decode the binary values of the group's character strings to the original values and then to learn and correct the neural network composed of the parameters.

Equation (25) below evaluates the performance of the neural network. The fitness of the neural network is calculated as the inverse of the sum of the loads of the learning and evaluation errors of the neural network, as shown in Equation (26).

The third step is an inspection step of the termination condition, and ends when the goodness of fit is above the target value or the number of households exceeds the number of designated households.

The fourth stage is the creation of a new generation, where the new generation creates new populations by breeding and mutating through regeneration, selection, and the population of the previous generation (see FIG. 8). The mating is performed according to the mating rate initially selected by selecting two objects according to the selection probability of Equation (28) below. Mutations change random bits by an initially determined mutation rate in a new generation of crossed chromosomes. Step 2 is performed when the generation of the new generation is complete.

Genetic algorithms using binary string coding do not guarantee the final convergence due to the limitations of precision of strings. Therefore, in the present invention, a set of superior objects of a genetic algorithm is regarded as an initial solution of a complex algorithm. Find the optimal value.

Neural network model predictive control performs predictive control based on output prediction and moving interval principle using neural network model. The purpose of the predictive control is to obtain a future control input for minimizing the cost function as shown in Eq. (29) for the target output trajectory of the prediction section and output only the first control input.

Constrained nonlinear optimization algorithm SQP (Successive Quadratic Programming) is used to generate the optimal control input. 9 shows a structural diagram of a nonlinear model predictive control using a neural network model and an SQP algorithm.

On the other hand, neural network model for daily water demand prediction is established through on-site data collected through actual operation in water purification plant and various factor data affecting water usage.

The neural network model is trained with the error backpropagation algorithm based on the periodically collected operational data, and finds the optimal prediction demand from the learned neural network model. On-site data indicating the water usage at the point to be predicted is transmitted to the host computer at each collection cycle through the LAN network, and the collected data is input to the neural network model to calculate and output the current demand demand at each calculation cycle. 10 shows the flow of the overall daily water demand forecasting method.

In addition, demand forecasting can be applied to various algorithms, but in practice, selection of exponential factors (factors) that affect water demand plays an important role in minimizing prediction errors. Generally, variables affecting water usage include rainfall (or precipitation), weather (cloud volume), day of the week, humidity, minimum temperature, maximum temperature, and average temperature, among which independent variables are used as input variables. In order to increase the reliability of the survey, 24 skilled professionals from the head office, research institute, training institute, and Cheongjuwon management team of the Korea Water Resources Corporation were surveyed. The results are shown in FIG.

In addition, in the short-term forecasting of tap water demand using Transfer Function model, a research paper related to demand forecasting (Jae-Joon Lee, Journal of Korean Water and Sewerage Society, pp88-103, 1996.3), the highest temperature, rainfall, day of the week, humidity, average temperature, weather and minimum temperature It is regarded as an influence factor of tap water use.

Based on the above, in the present invention, three independent variables, day of week (day-type), maximum temperature, and rainfall (precipitation) are selected as having a relatively high degree of relevance. Since we can assume that the accuracy of the prediction results is inherent, we select these four variables as inputs to the prediction process and model water demand.

The collection of each variable was carried out as follows to provide predictions that maintain accuracy above the expected level from the appropriate number of sample data.

1) Days of the week: Mondays to Saturdays (1 ~ 6), holidays (0) such as Sundays or public holidays, New Year's Day and Chuseok (1)

2) The highest temperature: The weather data is provided from the Korea Water Resources Corporation's weather database where the weather forecasting data is collected and stored, and the unit is ℃.

3) Rainfall: The Korea Water Resources Corporation's weather database does not provide a forecast of rainfall, so only precipitation is known, so precipitation is collected separately instead of rainfall. That is, it is divided into two parts, sunny / cloudy (0) and rain / snow (1).

Therefore, in the present invention, the daily demand demand prediction model is modeled using a neural network, and an example of input data is shown in the following table (1).

        division  date Day of the week Precipitation Temperature Daily demand Actual demand … … … … … … Sep 6, 2003 6 0 28 33682 35089 September 7, 2003 0 One 25 35089 34954 Sep 8, 2003 One One 27 34954 34239 9/9/2003 2 One 24 34239 32525 Sep 10, 2003 -One 0 26 32525 34136 September 11, 2003 -One 0 26 34136 29800 Sep 12, 2003 -One One 24 29800 30519 Sep 13, 2003 6 One 25 30519 32829 Sep 14, 2003 0 0 27 32829 35928 … … … … … …

In the present invention, for an effective neural network modeling, the neural network is composed of four input neurons, ten or more intermediate neurons, and one output neuron. As shown in FIG. 12, the neural network model is constructed by assigning the maximum temperature, daily format (day of the week), precipitation presence, and full-day performance as input variables and assigning water prediction usage as an output variable of the neural network model.

Genetic and complex algorithms are used to determine the optimal value of the neural network structure (number of neurons in the middle layer), learning parameters (learning rate, momentum coefficient, range of initial connection load, learning tolerance, and slope of sigmoid function). . FIG. 13 shows a string for searching for an optimal structure of a neural network structure and learning parameters.

As shown in Table 2, the character string of each parameter has a minimum value and a maximum value for decoding, and Table 3 shows the set values of parameters required for the execution of the genetic algorithm.

parameter  range Number of neurons in the hidden layer 1 to 30 Learning rate 0.0001 to 1.0 Momentum coefficient 0.0 to 1.0 Initial connection load range ㅁ 0.05 ~ ㅁ 2 Learning tolerance 0.001 to 0.05 Sigmoid Function Slope 0.0 to 1.0

parameter Set value Evolutionary generation number 500 Group size 100 Total string length 50 Mating rate 0.7 Mutation rate 0.01

In order to verify the usefulness of the proposed technique, a simulation was conducted based on the data collected at a water treatment plant of a regional water treatment system. The plant has an average processing capacity of 346,000㎥ and a digital distributed control system, and the quarter of the simulation data shows an average daily water consumption of about 33,000㎥.

The neural network model for daily demand forecasting was constructed by learning the data for five months of collectible data of the target quarter. The parameters of the neural network used are shown in the following table (4).

parameter  range Hidden Layer Nodes 23 Connected load learning rate 0.002264, Threshold Learning Rate 0.001837 Sigmoid Slope 0.070452 moment 0.004181 Differential moment 0.004727

Hereinafter, an embodiment of the demand forecasting software according to the present invention will be described with a computer screen.

The module on which the demand forecast is executed is built into the quantity management software. The quantity management software performs daily demand forecast by water supply area, and based on this, calculates daily demand forecast by 2 days (48 hours) by drainage pattern analysis as a basic function. The calculated demand forecast by hour calculates the amount of water supply by water supply area and establishes a two-day operation plan in order to economically operate in consideration of electricity rate, flow rate of each water supply area, and inflow smoothing.

If you look at this process step by step, after performing the daily demand forecast applying the results of this study, seasonal information (spring / autumn, summer, winter), weather information (sunny / cloudy, rain / snow), date information (weekdays, Find the demand pattern (time coefficient) from the demand pattern table. By multiplying the forecasted demand by the hour by the hour, the forecasted demand by hour is calculated, and the time coefficient of the demand pattern is calculated through the function of the demand forecasting model based on the performance value.

The demand pattern analysis performs the function of analyzing the demand pattern over time for the demand pattern which is representative among the 13 demand patterns. In order to produce reliable results when demand pattern analysis is executed, at least 12 months of historical data should be collected / stored, and the demand for demand and collected weather information are classified by date and season and used for demand pattern simulation. . However, new demand pattern analysis simulation is executed once a day online to facilitate the convenience of the operator and to reflect as much recent performance data as possible.

Finally, based on the hourly forecasted demand of 48 hours per system, the total pump operation number plan for each hour is automatically generated using the combination of pump operation number per demand. In other words, the pump operation number is planned for the demand by using the demand quantity and the pump performance curve (efficiency). The plan is based on the electricity tariff.

In order to provide the possibility of data collection and to provide predictions that maintain accuracy beyond the expected level from an appropriate number of sample data in order to reduce the associated time and cost, the data is classified by applying the data granulization technique, and demand by the converted average of the classified data Set the pattern. Data granulization is basically based on time coefficient and climate coefficient. Time coefficient is divided into season parameter and day-type parameter, and climate coefficient is weather parameter. Separated by. The details of each coefficient are as follows.

■ Seasonal Coefficient (SP): Spring / Autumn, Summer, Winter

■ Day Format Factor (DP): Weekdays, Holidays, Unusual Days

■ Weather Coefficient (WP): Sunny / Cloudy, Rain / Snow

However, it does not distinguish between seasonal and day-type coefficients, considering all 365 days of the year as time coefficients, and the items of weather coefficients are sunny, cloudy, and rain. The accuracy of each granule can be increased more by reflecting various cases such as snow, cloudy after cloudy, sunny after cloudy, etc. However, when all the cases are reflected, only a small number of pattern data data corresponding to a certain granule exist or do not exist at all. In some cases, and as mentioned above, if the accuracy of the expected data can be maintained from the appropriate number of sample data, the representativeness of each granule will not be significantly affected.

On the other hand, this method is to classify the time-variable data into classes according to characteristics, and constitutes various types of granules according to seasons, days of the week, and time, and classifies the granules into similar classes. By using this class data as historical data of the model function, effective pattern setting is possible.

The modeling method starts from different concepts by recognizing the large difference between socioeconomic parameters and other time-varying parameters in the approach. While socio-economic parameters do not allow frequent collection of data, time-variable parameters tend to collect statistical data due to their frequent changes. There are a total of 8760 data per year (24 hours 365 days), which can be classified into the same type of data by seasonal and weekday characteristics. For example, there will be several data that are summer, weekday, and the same time zone. In this way, 8760 total data are classified into the same type of data, and these data are historical data that estimate the demand pattern of the characteristic class in each situation. Used as The requirements for this are as follows. Class classification (drainage pattern classification) is shown in Table 5.

■ Collect historical data required for every granule over time

One year data collection: 24 (hours) × 365 (day) = 8760 ()

■ Total number of classes for SP, DP, WP (see Table 6)

SP (3) × DP (2) × WP (2) + 1 = 13 (class)

season Japanese style weather Unusual spring fall summer winter weekday holiday Sunny / Cloudy Rain / snow One ◎ ◎ ◎ 2 ◎ ◎ ◎ 3 ◎ ◎ ◎ 4 ◎ ◎ ◎ 5 ◎ ◎ ◎ 6 ◎ ◎ ◎ 7 ◎ ◎ ◎ 8 ◎ ◎ ◎ 9 ◎ ◎ ◎ 10 ◎ ◎ ◎ 11 ◎ ◎ ◎ 12 ◎ ◎ ◎ 13 ◎

The initial screen of the quantity management software is shown in FIG. 22, and the target individual setting is selected in the toolbar on the left. Neural network editing and learning are performed at the same time because the neural network model is constructed and learned according to each object. Up to 20 targets can be used, which means that it is possible to forecast water demand for 20 target quarters simultaneously during online operations. Editing of the object consists of neural network learning, input tag, output tag, and detailed information so that the target technology and input / output tag setting part can be done. I can move it.

To register a target for prediction, input a target name and press a tag designation button on an input tag and output tag setting screen to register a tag corresponding to a desired variable as shown in FIG.

If I / O tag is selected, training data can be input manually or collected online according to the set period of the variable set, and basic setting for constructing neural network model is completed. Screens in which the setting is completed are shown in FIGS. 27 and 28.

The detailed information is the setting part corresponding to the case where the pump switching table is generated based on the water demand calculated through the hourly demand forecast after most of the daily demand forecast is as shown in FIG. 29.

The part that needs to be set during the forecasting of daily demand is the setting of the forecast area and the AWS location.In order to read the precipitation and maximum temperature, which are used as input data for neural network learning, from the Korea Water Resources Corporation's weather database, You need to set the branch code of AWS (Automatic Weather Station) which observes forecast area and actual wind direction, wind speed, temperature, and rainfall.

Pressing the forecast zone setting inquiry button or AWS site setting inquiry button in the detailed information setting screen can register the forecast area (FIG. 30) and the AWS branch (FIG. 31) where the target is located as follows.

In order to perform neural network modeling and learning, learning data should be secured first. By editing only the target information, online water quality data can be collected during operation, and existing weather data and performance demand can be edited and used as learning data. The editing of learning data is discussed later. When learning data based on target information is constructed, neural network modeling and built modeling are learned and used for online operation. In FIG. 23, the user may manually input the learning parameter and the learning goal, but may be automatically generated through the optimum model search function.

Table 6 shows the range and each parameter that can be specified when setting learning parameters.

Item Contents Remarks  RMS target error  0.0001 to 15.0 If the RMS error, the least-squares error that is the performance goal of the model, is less than or equal to the target error, the learning ends.  Number of data pairs  1 to 32767 The number of data pairs to learn  Learning frequency  1 to 1000000 Model training count  Hidden Layer Nodes  1 to 30 Set value close to 30  Connection load learning rate  0.0001 to 1.0 Usually set to a value less than 0.05  Threshold learning rate  0.0001 to 1.0 Usually set to a value less than 0.05  Sigmoid Function Slope  0.0001 to 5.0 Usually set to a value between 1.0 and 5.0  moment  0.0001 to 1.0 Usually set to a value less than 0.01  Differential moment  0.0001 to 1.0 Usually set to a value less than 0.01  Training data file name  Up to 20 characters in English or Korean Extension is ".csv", format is [number of data] [number of input] [number of output] [data] and can be edited with Excel  Connection load file name  Up to 20 characters in English or Korean Extension is ".wgt", format is [number of inputs] [number of hidden layer nodes] [number of outputs] [connection load]  Test data file name  Up to 20 characters in English or Korean Extension ".csv" format can be edited by [Number of data] [Number of input] [Number of output] [Data], Excel

If you press the button to search for the best model in the lower left, as shown in Figure 24 automatically generates the optimal neural network learning parameters.

Once the training data and neural network modeling are established, offline learning must be performed. The learning is stopped after satisfying the RMS target error set as the learning goal or after learning as many times as the learning goal and storing the result in the link load file. The stored learning load is used to calculate the control output according to the values of the input variables during online operation. The initial load file is used to perform the previous learning back to the learning goal. 25 is a screen on which neural network learning is executed.

Once you have learned, you can review the learning results through learning assessment. This is to compare the difference between the model output and the value of the output variable of the training data using the training data. If you press the learning evaluation button in the above edit screen, you can review the learning results as shown in FIG.

The test evaluation function is similar to the learning evaluation function, and is used to compare the learning results with the untrained data.

FIG. 33 is an edit screen of learning data displayed when selecting learning data from a toolbar of an initial screen, and is a function of manually editing data by opening data collected online or existing learning data based on neural network file information. .

One learning file (learning data file set in neural network learning) is assigned to one prediction object. If the target number item is changed, the learning file is also moved to the corresponding file. The file name is displayed at the top of the screen, and it can also be collected online only if it is identical to the learning data file of the operation schedule set up later.

The screen displays the number of inputs and outputs and distinguishes the color of the data corresponding to the input data and the output so that the user can easily recognize them. This data is saved in csv extension format and is compatible with MS excel program. If you are familiar with MS Excel program, you can use it to edit.

If the operation screen is selected in the toolbar of the initial screen, each target information and the set operation information are displayed as shown in FIG. 34.

On-line operation can perform daily pattern collection / pattern analysis execution, demand forecast execution, pump transfer table generation, and training data collection function. These functions are performed by the set time (cycle). The operation and time (period) of each function can be set by selecting the operation schedule setting button at the bottom of the operation screen.The operation schedule is recorded in the registry and set only once.After setting the operation schedule, press the operation start button to operate online. Just start The functions corresponding to the daily demand forecast related to the present study are demand forecast execution and learning data collection as shown in FIGS. 35 and 36, respectively.

The operation start screen shows operation status information as shown in FIG.

Although the preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the above-described embodiments, and the present invention is not limited to the above-described embodiments without departing from the spirit of the present invention as claimed in the claims. Various modifications can be made by those skilled in the art, and such modifications are intended to fall within the scope of the appended claims.

As described above, according to the water demand prediction method using the neural network according to the present invention, there is an effect of determining the optimal prediction demand amount using intelligent modeling and nonlinear model predictive control technique for stable operation of the water operation system.

In addition, the average error between the forecasted demand and the actual demand of the intelligent model should be set to 10% or less, and the past operational data of the target quarters, which are indirect information, are selected as the criteria for demand forecasting and stabilization targets. By performing the operation by online setting, it is possible to reduce the manpower and the efficiency of manual operation by automation of demand forecasting.

In addition, in the present invention, the optimal prediction demand is determined by neural network modeling and nonlinear model predictive control technique of daily demand forecasting process, and nonlinear process optimization and auto tuning of a manual operation system where high efficiency and stability of the demand forecasting process are essential. It implements the optimal structure and parameter search of the manual operation software and helps to improve the reliability and maintenance of the system, standardization, and secure technology, and also improve the overall efficiency and function of the manual operation system by the advanced control technology. There is an effect that can be achieved.

1 is a view showing an overview of a general water operating system.

2 is an operational flow diagram for the main functions of one demand forecast.

3 is a conceptual diagram of a Kalman filter model sequentially estimating the state of a system;

4 shows a discrete dynamic linear system of a Kalman filter.

5 shows the neural network structure of a single neuron.

6 is a diagram showing a neural network model having a multilayer structure.

7 shows a string with a binary code;

8 is a view for explaining the process of generating a new population by breeding and mutation through reproduction, selection of the previous generation of individuals.

9 is a structural diagram of neural network model-based predictive control.

10 is a flowchart illustrating a method of setting daily water demand forecasting.

11 is a graph showing a questionnaire result of an input variable to be used in demand forecasting.

12 is a diagram showing a neural network model of water demand prediction.

13 is a string for searching the structure and learning parameters of the neural network.

14 is a hardware structure diagram for automatic calculation of daily water demand demand.

15 is a software operation flowchart for automatic calculation of daily water demand demand.

16 is a graph showing the test target prediction results of spring and autumn.

17 is a graph showing the prediction result of the test subject in summer.

18 is a graph showing a test target prediction result in winter.

19 is a graph showing a test subject prediction error in spring and autumn.

20 is a graph showing a test subject prediction error in summer.

21 is a graph showing a test subject prediction error in winter.

Fig. 22 is a water management software initial screen;

23 to 32 are screens showing the setting of learning parameters and learning goals.

33 to 36 are screens showing learning data editing and online demand forecasting.

37 is a screen illustrating evaluation of learning results.

38 is a screen showing demand forecast online operation.

Claims (5)

In the usual water demand prediction method, Water demand forecasting system using neural networks and evolutionary algorithms to secure stable supply by reasonably linking intake, water purification and pressurization stations reflecting the usual changes due to demand fluctuations, weather conditions, seasonal factors, etc. Building an intelligent model of the; Configuring demand prediction software on a higher level computer that calculates the demand for the predicted target by using a neural network model-based predictive control technique based on the intelligent model; Connecting to a host computer of a manual operating system, a database having weather information, and the like through a LAN network; The demand forecasting software may include constructing a demand forecast neural network model using a backpropagation learning algorithm based on operational data collected from a system and data of a weather DB; The neural network model-based predictive control algorithm from the intelligent model using the neural network, characterized in that the step of determining the current demand and forecast demand in weather conditions. The method of claim 1, For stable operation of the manual operation system, the optimal forecasting demand is determined by using intelligent modeling and nonlinear model predictive control technique, but the optimal demand forecasting model is built using intelligent modeling techniques such as neural network and evolutionary algorithm in addition to the host computer. And a process demand prediction method using a neural network comprising a process simulator configured to determine an optimal demand for demand using a nonlinear model predictive control method based on the intelligent process model. The method of claim 1, The neural network model is a water demand prediction method using a neural network, characterized in that for applying a back-propagation (BP) neural network model. The method of claim 1, The basic structure of the neural network is composed of an input layer, an intermediate layer, and an output layer. Each layer is composed of nodes corresponding to neurons of a neural tissue, and each layer is connected to each other by a link weight. A method for predicting water demand using a neural network, which uses a sigmoid-type function having an activation function or a logistic function. The method of claim 1, The cost function is composed of the norm of the output error, but by adding the norm of the error rate of change, the water demand prediction method using the neural network improves the performance of the overall algorithm by adding the learning performance in the direction of minimizing the error rate of change. .