JPH03134703A - Demand pattern predicting method and plant operating method - Google Patents

Abstract

PURPOSE:To execute an operation adapted to an actual result performed by an operator and the previous instance without necessitating much labor by inputting the predicted value of a disturbance factor on that day to a learned neural net and predicting a demand pattern on that day. CONSTITUTION:A leaning neural net 71 is allowed to learn a disturbance factor and an actual result demand pattern at plural time points in the past. The constitution of the neural net 71 is a hierarchical structure consisting of an input layer 710 constituted, based on a neuron model as a foundation, an intermediate layer 720 of at least one layer, an output layer 730, a comparison layer 740, and a teacher layer 750, and learning is executed by an error reverse propagation method. Subsequently, a disturbance factor predicted value on that day is inputted the input layer 710 of a learned neural net (predicting neural net) 72 and the output layer 730 is allowed to output a demand pattern on that day and it is predicted. In such a manner, the operation adapted to an actual result performed by an operator and the previous instance can be executed without executing a large quantity of data analysis and an interview to the operator.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention provides a method for predicting demand patterns on a daily or weekly basis, such as water demand and the amount of water flowing into a sewage treatment plant, and a method based on the results of these predictions. The present invention relates to a method for controlling the operation of a supply plant.

[Conventional technology]

Demand patterns for water demand and the amount of water flowing into sewage treatment plants fluctuate on a daily or weekly basis, reflecting human social activities. For example, in a day, demand is high during the day, and conversely, it is low at night.

These predictions are important because water supply plants and the like use these predicted values to make operational plans for supply amounts in advance. These demand patterns are not only strongly influenced by uncertain factors such as the day's weather and temperature, but also have the characteristic that the influence of the previous day's actual pattern remains on the day (1
! history). Conventional methods include statistical methods such as autoregressive models, and predictive methods based on knowledge engineering that utilize the experience of supply plant operators.

[Problem to be solved by the invention]

In order to use conventional statistical methods, it is necessary to organize and analyze a large amount of actual data in advance. On the other hand, the method of applying knowledge engineering requires interviews with operators to create, modify, improve, and add knowledge rules, and to organize empirical knowledge without contradiction. As described above, the conventional method required a great deal of labor. In addition, in the knowledge engineering method,
If contradictory conclusions are reached at the same time, no final conclusion can be drawn)
There were drawbacks.

[Means to solve the problem]

The present invention is characterized by using a neural network (neural circuit model) to predict the demand turn based on past performance.The operator can determine the degree to which disturbance factors such as weather and temperature affect the demand pattern, and Demand forecasting is done empirically, taking into account historical characteristics.In the operator's brain, predictions are made based on these "past results," and this action is realized using a neural network.Here, , "Past performance" is a typical demand pattern at multiple points in the past.

Therefore, first, the neural network is made to learn disturbance factors and actual demand patterns at multiple points in the past. The specific configuration of the neural network that executes this is to have at least one input layer based on a neuron element model.
It has a hierarchical structure consisting of an intermediate layer, an output layer, a comparison layer, and a teacher layer.

A specific learning method is a known error backpropagation method. Next, the predicted value of the disturbance factor for the day was input into the input layer of the trained neural network, and the output layer outputted the demand pattern for the day. This action is prediction. In addition, learning can be performed automatically or arbitrarily by the operator, so that predictions can be made accurately at all times. Furthermore, the predicted demand pattern is reflected in the operation method, such as by formulating a supply operation plan and controlling the target value of supply amount.

[Effect]

In the present invention, the learning prediction ability of a neural network is applied to predicting demand patterns. That is, first, the neural network is made to learn the history of disturbance factors and demand patterns. At this time, the distribution of weighting coefficients of the neural network changes. This corresponds to the operator memorizing various past achievements. By inputting predicted values of disturbance factors for the day into such a trained neural network, the demand pattern for the day can be predicted. This means that
This simulates the method by which an operator predicts demand patterns by taking into account various past results.

〔Example〕

The present invention allows a neural network to learn past disturbance factors and actual demand patterns on a plurality of sides, and inputs the current day's disturbance factors into the learned neural network to associate the same day's demand pattern. The execution process consists of (1) a learning process using a learning neural network model (learning neural network), (2) a demand pattern prediction process by association of a trained neural network (prediction neural network), and (3) a process based on the prediction results. It consists of an operation control process based on the supply amount.

Applications include predicting water demand patterns at water treatment plants and the amount of sewage flowing into sewage treatment plants. Here, the configuration and operation will be explained using the embodiment shown in FIG. 1, taking as an example the water demand pattern prediction of a water purification plant.

First, the configuration of the first cap body will be explained. In Figure 1,
Chlorine-injected purified water produced at a water treatment plant is stored in a water distribution reservoir 21. Purified water in the distribution reservoir 21 is sent to the pump 22 through a pipe 21P1, and then distributed to consumers through a pipe 21P2. Although not shown, a valve is installed in the pipe 21P2 to control pressure and flow rate. The time series data of the flow rate, which is the operation data of the pump 22, is sequentially stored in the history pattern data file 71F in the pre-ignition operation control device 80. The predictive operation control device 8o is a computer system, and in order to facilitate the explanation of the present invention, the first
The figure shows a flow diagram of the process.

In addition, the portion described as network in the figure corresponds to the solid line 702 in FIG. 4.

The time series of the flow rate signal 22Q shows the integrated amount per hour as 22Q.
0 and the value Q (t, i) (where t: day Hl a
time). On the other hand, the operator 101 communicates information such as the weather, day of the week, average temperature, maximum temperature, temperature at representative times, amount of solar radiation, presence or absence of an event and its degree, etc. through a communication means 46 consisting of a display, keyboard, or voice input/output device. As the disturbance signal 46G, the predictive operation control device 8
0 to the history pattern data file 71F. Furthermore, data signals such as the weather information can be inputted to the communication means 46 from the outside through the communication line 46I.

Note that when the disturbance signal 46G is measured by an automatic measuring instrument, it is of course automatically input. In this way,
The flow rate signal 22Q and the disturbance signal 46G are sequentially stored in the history pattern data file 71F.

The flow rate signal 2 is automatically generated from the history pattern data file 71F or by an instruction 71S from the communication means 46.
The selected data string Q of 2Q and the selected data string G of the disturbance signal 46G are each output to the learning neural network 71. Learning is executed using these data Q and G, and execution results 71S1 and 71S2 are output to the prediction neural network 72. Prediction neural network 72
Then, receive the execution results 71S1 and 71S2.

The data string of the flow rate signal 22Q and the data string of the disturbance signal 46G necessary for prediction are selected from the history pattern data file 71F and used for the prediction.

The prediction result is displayed on the communication means 46 as a prediction signal 72S, and is also output to the operation control process 75.

The operation control process 75 receives the prediction signal 72S and outputs a flow rate target value signal 75S to the pump 22 to control the flow rate.

At the same time, the flow rate signal 75S of the target value is displayed on the communication means 46, and the actual flow rate is corrected as required by the operator 101's selection. The correction value is again output to the pump 22 as a flow rate signal 75S.

Next, the present invention will be explained in detail using FIG.

At the same time, FIG. 2 and subsequent figures will be used to explain the detailed configuration and operation.

First, a method of storing historical pattern data in the historical pattern data file 71F will be explained. The history pattern data consists of a flow rate signal 22Q and a disturbance signal 46G.

The flow rate signal 22Q of the pump 22 is expressed as Q(t,i) (where,
is two days, j: time), then data Q (0, 1) from 1 o'clock to 24 o'clock (j = 1 to 24) on that day (1 = 0),
Q(0,2), Q(0,3), . . . Q(0,24) are as shown in FIG. These data are output to the history pattern data file 71F.

If this is repeated every day, data of t=o, -1, -2, . . . will be stored sequentially. On the other hand, if the disturbance signal 46G input through the communication means 46 is G(tlj) (here, t: day, j: disturbance number), on the current day (1=0)
weather = G (0, 1), day of the week = G (0, 2), average temperature = G (0, 3>, maximum temperature = G (0, 4), temperature at representative time = G (0, 5) ), solar radiation = G (0,6)
, presence or absence of event = G (0, 7) and degree = G (0, 8)
etc. are converted into numerical values as shown in Figure 2. These data are output to the history pattern data file 71F.

Next, the learning neural network 71 in the learning process
The operation will be explained below. Learning neural network 71
Now, data selected from the history pattern data file 71F is received. This data selection method will be explained below using FIGS. 1 and 3. Disturbance signal G(t, j
), the block of the learning neural network 71 in Figure 1 and the selected time (day) 1 as shown in Figure 3.
=11°tzttat..., go back from t and use an arbitrary number of days. Day 1 (1=1 work#t
2jt31...), G(t,
1), . . . +a (t+a) are input, and for the previous day, G (t-1, 1), . . . , G (t-1, 8) are input. However, weather G (0, 1), day of the week G (0°2
) The presence or absence of an event G(0,7) and the degree of the event G(0
, 8) are input by converting them into symbols or numbers. On the other hand, for the flow rate Q(t, i), 1=1
Flow rate Q for 24 hours at □+tztt□,...)
(t, 1).

..., Q(t, 24) are input to the teacher signal layer 750. On the other hand, for the input layer 710, the value Q(t-1,1) of the previous day (t=-1).

..., Q (t-1, 24) and the values before the previous day are entered in the past for an arbitrary number of days. If the number of days ρ is large, the prediction accuracy will improve, but in practical terms, the previous day (t=
-1) and data from the day before (t=-2) are sufficient. In addition, multiple points in time consisting of nine past points are 1 = 1 t 2 1 t 3
t...t, pattern data to be learned is input at time t, (i=1.2...) as shown in FIG.

For each data of the input layer 710 set in this way and the teacher layer 750, intermediate/! ! 720, output/1
1730, learning is performed by comparison layer 740. The signal processing method in this learning method will be explained below using FIG. 4. The configuration and signal processing method shown in FIG. 4 are well known, except for the input data setting method described above. That is,
The configuration and signal processing method shown in FIG. 4 are publicly known techniques devised by Rumelhart et al.
rallell DistributedProces
sing, MIT Press, Vol. 1. (19
86)).

The configuration and operation of FIG. 4 will be outlined. First, the symbols in FIG. 4 will be explained. O is a neuron element model 701, and a real m702 connecting O and O indicates that information is exchanged between the neuron element models 701. Here, each layer consists of a finite number of neuron element models, and all neuron element models in adjacent layers are connected. Although there may be a plurality of intermediate layers 720, this embodiment shows an example in which the number of intermediate layers is one for ease of explanation. The configuration shown in FIG. 4 is called a neural network (neural circuit model).

Next, basic operations of the neuron element model 701 will be explained using FIG. As mentioned above, the data input to the input layer 710 in FIG. 4 is G(tyj)(t=
tyt 1. ...; j=1+..., 8) and Q
(Li) (t=t-1, t-2,...; x=1.
..., 24), but for general explanation -G
(t, j) and Q(Li) (assuming that there are n in total), the n variable values are expressed as Yi~ as shown in Figure 5.
Let it be Yn. An operation (sum-of-products operation) in which each of the input signal values Y□ to Yn is multiplied by a weighting coefficient Wji and then added together is calculated using equation (1).

Here, Yi (1): the value of Yi of the input layer (first layer), W
ji (2←1) Weighting coefficient from the i-th variable of the two-person layer (1st W) to the j-th three-neelon element model of the intermediate layer (2N), Zj (2): is the input summation value to the j-th neuron element model of layer).

In the neuron element model 701, the output value here is calculated using equation (2) depending on the magnitude of Zj(2).

The calculation details of equation (2) are as shown in FIG.

The calculated value Yj(2) is further sent to the output layer, and the same calculation is performed in the output layer.

Next, an overview of the calculation method using a neural network will be explained. The variable value Yi(1) mentioned above is input to the input layer 7 in FIG.
10, and this signal value is output to the neuron element model of the hidden layer 720. In the neuron element model of the intermediate layer 720, these output values Yi(1) and weighting coefficients Wi
Calculate the sum of products Zj (2) with j (2←1) using formula (1),
Depending on the magnitude, the output value Yj(2) to be output to the output layer 730 is determined using equation (2). Similarly, the output value Yj(2) of the intermediate layer 720 is further calculated as the sum of products Zj(3 ) is calculated using equation (3).

Here, Yi(2): value of the middle layer (second layer), Wji(
3←2): Weighting coefficient from the i-th variable of the intermediate layer (2Nth layer) to the j-th neuron element model of the output layer (3rd layer), Zj (3): j of the output layer (3rd layer) is the total input value to the th neuron element model.

Furthermore, the output value Yj(3) to be outputted to the output M730 is calculated using equation (4) depending on the magnitude of Zj(3).

In this way, the calculated value Yj(3) of the output layer is obtained. In Figure 1, the pattern of Yj(3) is Q(t, 1)*1
Q(t, 2) books, . . . IQ(tT24) books.

To perform learning with a neural network, the output layer 73
After 0, a comparison layer 740 and a teacher signal layer 750 are further provided, and the signal 730s of the output layer 730 and the teacher signal /l7
50 teacher signal 750S enters the comparison M740, where the output signal 730S Q (t, 1) * (however, Cl-24) and the teacher signal 750S Q (t, i) (
However, the size is compared with C1-24). The magnitudes of the weighting coefficients Wji (3←2) and Wji (2←1) are corrected so that this deviation becomes smaller. When the calculations of equations (1) to (4) are executed again using this corrected value and the magnitudes of the output signal 730S and the teacher signal 750S are compared, a deviation similarly appears. Based on this deviation, the magnitudes of the weighting coefficients <Vji (3←2) and Wji (2←1) are corrected again. In this way, the weighting coefficient Wji is repeatedly corrected until the deviation becomes equal to or less than a predetermined value. Initially, the weighting coefficients are randomly given by generating random numbers, so the deviation is naturally large, but the output signal value gradually approaches the teacher signal value. At this time, (1)
- In equation (4), the value that is changed is the weighting coefficient value Wj
Since there is only i, the learning result is reflected in the distribution of Wji values.

The method of correcting the deviation in this manner is called the error backpropagation method, and uses a known technique devised by Rumelhart et al. For details, see the literature (Parallel Di
distributed processing, M.I.T.
Press, Vol. 1. (1986)). Although this learning method itself is well known,
In particular, the present invention repeatedly learns a group of historical pattern data at different points in time, and the weighting coefficient w
It is accumulated and stored in the distribution of j. This allows the system to have the same effect as the operator's past experience.

Multiple time points 1=1□+t21t3? ... t, a learning method for the historical pattern data group will be described below. A plurality of historical pattern data groups are learned by arbitrarily changing the date t. This day t, 1t21t3t...t may be selected by the operator or automatically, and both will be explained below.

The operator's selection is a typical pattern of variable values Y~Yn that he/she would like to reflect in the operation later, or a pattern for an abnormality that he/she would like to refer to later. As a result, the neural network outputs results according to what it has learned, so this selection is important. Leaving the selection to the operator relies on the operator's empirical and comprehensive ability to judge data. In this case, the patterns to be learned are patterns on different days, and a plurality of patterns are repeatedly learned.

This gives the neural network the ability to grasp patterns comparable to the operator's past experience.

The operator sets the date t through man-machine conversation via the communication means 46. For example, 12 patterns are learned from a typical pattern for January to a typical pattern for December.

On the other hand, when it is performed automatically, statistical analysis of the data string is required in advance. That is, the case with the highest frequency of occurrence is determined by statistical analysis and is regarded as a representative example of normal conditions, and this is learned, while the case of low occurrence frequency is regarded as a typical example of abnormal conditions and these are learned.

The history pattern data group at multiple time points t"ti lt2 Ft3, . . . t, selected by the operator or automatically in this way is shown in FIG.

Next, the prediction process using the prediction neural network 72 (2) will be explained. The configuration of the prediction neural network 72 is shown in FIG. As shown in FIG. 1, the prediction neural network 72 predicts a demand pattern based on the learning results of the learning neural network 71, that is, the distribution of weighting coefficient values Wji. The values of weighting coefficients Wji (3←2) and Wji (2←1) obtained as a result of learning are output to the prediction neural network 72. Predictions are made using these. This method will be explained using FIG. 8. As shown in FIG. 8, first, as an input layer pattern in the input layer 710, a variable value Y is set based on the current day (t=0).
Input i (the predicted value of G for the current day, and the values of G and Q from the previous day and the day before the previous day) to the input I 710. Next, (1)-(4)
A demand pattern is output as an output layer pattern at the output N730 by calculation using the formula. What is the demand pattern?
Q (o, i) * (however, CI
-24). In this embodiment, this calculation is called "prediction." The calculated value by prediction depends on the learned pattern, so the pattern to be learned is a typical pattern at different times or a pattern at a noteworthy abnormality.
What must be done is as described above.

This prediction operation is performed before the start of the day's operation (usually the day before). In other words, the pattern of Q and G up to two days ago is 2
By inputting the number of days and the expected value of the disturbance factor for that day, the neural network predicts the demand pattern for that day.

In the operation control process 75, this demand pattern Q (0,i
) Upon receiving the signal 72S (Cl-24), a target flow rate signal 75S is output to the pump 22 to control the flow rate. Demand pattern Q (0, i)* is 1 in this example
Although the flow rate is for each hour, the time unit can be set arbitrarily. Of course, the smaller the time interval, the better the prediction accuracy.

If the set time interval (one hour in this embodiment) is too long to predict the flow rate over a short period of time (for example, one minute), the prediction is performed by mathematical interpolation.

At the same time, the flow rate signal 75S of the target value (prediction result) is displayed on the communication means 46, and the actual flow rate is corrected according to selection by the operator 101 as necessary. For example, the predicted flow rate is corrected by multiplying the proportional coefficient by the factor and adding 2 to the constant, and outputs this correction value again to the pump 22 as the flow rate signal 75S, and the pump 22 supplies water by this flow rate. do.

The effects of this embodiment will be described below. Conventional statistical forecasting methods and demand pattern forecasting methods using knowledge engineering required a lot of effort such as analyzing large amounts of data and interviewing operators, but by applying the present invention, operators can It is possible to carry out ``operation in line with actual results and precedents,'' which is carried out by the company. Furthermore, knowledge engineering can predict demand patterns even in cases where no conclusion can be drawn. Furthermore, since learning can be performed at any time, it is possible to quickly follow changes in the situation, learn and predict.

Next, another embodiment will be described using FIG. 9. This embodiment has the same configuration as the previously described embodiment shown in FIG. 1, except that the neural network configuration is divided for each day of the week. This is an example in which learning patterns are prepared for each day of the week, focusing on the fact that the demand pattern shows a characteristic pattern for each day of the week. As shown in Figure 8, the neural networks in the learning process are 7101 for Monday, 71D2 for Tuesday, and 71D for Wednesday.
D3, wooden marker day as 71D4, Friday as 71D5, Saturday as 71D6, and Sunday as 71D7. 71D
1-7107 respectively correspond to 71 in FIG. Similarly, in the prediction process, the 72DI of the prediction neural network
Prepare up to -72D7.

These neural networks perform learning and prediction for each day of the week. The operation control process 75 is similar to the embodiment of FIG. The embodiment shown in FIG. 9 has the effect of being able to accurately predict changes in demand patterns for each day of the week.

The embodiment shown in FIG. 10 is an embodiment in which the net is divided according to weather. The learning neural networks are 71T1, a neural network that learns sunny patterns, 71T2, a neural network that learns cloudy patterns, and rain (including snow).
The neural network that learns the pattern is configured as 71D3. 71T1-71T3 correspond to 71 in FIG. 1, respectively. Similarly, neural networks 72T1 to 72T3 are prepared as prediction neural networks. These neural networks perform learning and prediction for each weather type. The operation control process 75 is similar to the embodiment of FIG. The embodiment shown in FIG. 10 has the effect of accurately predicting changes in demand patterns depending on the weather.

An embodiment of the present invention has been described above, taking as an example the application to water demand pattern prediction and a water supply plant operating method. A similar method can be applied to prediction and operation control of plants that are operated according to past history and operation results, such as inflow prediction of sewage treatment plants. Specifically, in predicting the inflow sewage flow rate pattern of a sewage treatment plant, the inflow sewage flow rate and disturbance factors are stored in the history pattern data file 71F, and Q in the embodiment of FIG. 1 is the inflow sewage flow rate.

〔Effect of the invention〕

Conventional statistical forecasting methods and demand pattern forecasting methods using knowledge engineering require a lot of effort, such as analyzing large amounts of data and interviewing operators. By applying the present invention, it is possible to "operate in accordance with actual results and precedents" that operators are currently implementing, without analyzing a large amount of data or interviewing operators. Furthermore, knowledge engineering can predict demand patterns even in cases where no conclusion can be drawn. Furthermore, since learning can be performed at any time, changes in the situation can be quickly followed and predicted. Furthermore, since predictions can be made taking into consideration the characteristics of the demand pattern for each day of the week and the characteristics for each weather, it is possible to perform predictions with higher accuracy than conventional methods.

[Brief explanation of the drawing]

Fig. 1 is a configuration diagram showing an embodiment of the present invention, Fig. 2 is an explanatory diagram explaining a history data file, Fig. 3 is an explanatory diagram of a learning pattern, Fig. 4 is a configuration diagram of a neural network, and Fig. 5 is an explanatory diagram showing the neuron element model, Fig. 6 is a characteristic diagram showing signal conversion in the neuron element model, Figs. 7 and 8 are explanatory diagrams explaining the prediction process, and Figs. 9 and 10 are from this book. FIG. 3 is a configuration diagram showing another embodiment of the invention. 21... Water distribution reservoir, 22... Pump, 46... Communication means, 101... Operator, 71F... History pattern data file, 71... Learning process, 72...
Prediction process, 75... Operation control process, 710... Input layer, 720... Middle layer, 730... Output layer, 740
... Comparison layer, 750 ... Teacher layer, 701 ... Neuron element model. 2nd factor Figure 3 Figure 7 Figure 8 ±h Input layer Figure 5 Figure 4 Figure 9 Figure 7R Figure 6

Claims (1)

[Claims] 1. A method for predicting a corresponding demand pattern according to expected values of a plurality of disturbance variable values that change over time, comprising an input layer, at least one intermediate layer, and an output layer. The neural circuit model is equipped with a neural circuit model having a hierarchical structure, and inputs a plurality of different actual demand patterns and a plurality of actual disturbance variable values from the past of the plant, and uses the actual demand pattern corresponding to the input as a teacher pattern. A demand pattern prediction method comprising: learning a neural circuit model; and inputting expected values of a plurality of unlearned disturbance variable values into the learned neural circuit model to predict a corresponding demand pattern. 2. A demand pattern prediction method, wherein the demand pattern predicted in claim 1 is a water demand pattern of a water supply plant. 3. A demand pattern prediction method, wherein the demand pattern predicted in claim 1 is an inflow water flow rate pattern of a sewage treatment plant. 4. A plant operating method according to any one of claims 1 to 3, characterized in that the operation of a water supply plant or a sewage treatment plant is controlled using the predicted demand pattern as a target pattern.