JP2003030621A - Generated hydraulic power prediction method for run-of- river type dam and neural network therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly accurately predict the generated power of a run-of-river type dam. SOLUTION: In the power plant of the run-of-river type dam, since a water storage function for preventing the flood of a downstream area is not originally provided differently from a normal dam, flow rate prediction for flood prevention is not required to begin with. Thus, without tentatively predicting a flow rate, the generated power is directly predicted. That is, the output of a neural network is defined as power generator output (power generator output difference, in this example).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of predicting hydroelectric power generation of a self-propelled dam using a neural network.

[0002]

2. Description of the Related Art Prediction of the amount of hydroelectric power generation and the amount of inflow in a dam is made from the viewpoint of economy and safety. This is because if the flow rate that will flow into the future can be accurately predicted, not only will it be possible to reduce the amount of ineffective discharge and increase the amount of power generated, but it will also be possible to prevent floods in the downstream region during heavy rainfall.

As a method of predicting the amount of power generation, there are a method of directly predicting the amount of power generation and a method of predicting the inflow amount and then converting the predicted inflow amount into the amount of power generation for prediction. Only the latter method is used. The reason is that the operation is often focused on safety, and the main purpose is to prevent flooding in the downstream region during heavy rainfall. Therefore, the prediction of the power generation amount is performed by applying the predicted inflow amount to the QP characteristic (power generation flow rate-power generation amount characteristic) of the hydroelectric generator, rather than directly predicting the power generation amount. This means that the prediction accuracy of the inflow amount is directly connected to the prediction accuracy of the power generation amount.

The power generation flow rate is a flow rate that flows into the generator during the inflow amount (when the inflow amount is large, a part of the flow amount is directly discharged). Generally, the inflow is predicted by using the rainfall in the upstream area, the flow rate of the upstream river, and the actual value of the inflow to the dam in the past several hours. There is a very strong non-linearity between future inflows and these data, making them difficult to predict. For example, even if the amount of rainfall is the same, the flow rate varies greatly depending on the wetness of the ground. Moreover, the flow velocity of the river becomes very fast when the flow rate is high.

Since it is very difficult to predict the flow rate, it is difficult to accurately predict the flow rate by using a general regression equation. Normally, the tank model or the storage function method is used, and adjustment is performed so that the prediction accuracy is best when the flow rate reaches a peak when the flow rate starts to increase during rainfall (where the flow rate is relatively high). (Because it is not possible to predict with high accuracy for all data from a small amount to a large amount). In recent years,
A prediction method using a neural network has also been put into practical use. A neural network has attracted attention as a method with relatively low prediction cost and relatively high prediction accuracy because a prediction model can be constructed relatively easily by simply giving data and learning.

[0006]

Here, the self-flow dam is a dam that has no gate for storing water and flows the inflowing water as it is downstream. Since discharge control is not possible, it is usually constructed at the uppermost stream position where flood protection in the downstream area is low.

Prediction of hydroelectric power generation in a self-propelled dam is
It is very effective from the economical point of view. For example, if the prediction accuracy of the power generation amount is high, it is possible to save the energy consumption amount such as thermal power generation.

The size of the hydroelectric generator installed in the self-propelled dam is generally determined according to the flow rate of the river.
In other words, introducing too large a generator in a river with a low flow is wasteful in terms of installation cost, and conversely introducing a small generator in a river with a large flow is a waste of water resources. There will be many. Generally, hydropower generators are designed to provide maximum output during a fraction of the year (in rivers in Japan, the output of hydropower generators can easily be reduced even when rainfall does not reach flood discharge). Maximum).

In predicting the amount of hydroelectric power generation in a self-propelled dam, the prediction accuracy when the amount of inflow is comparatively small (the range in which the amount of power generation changes according to the flow rate before the generator saturates at maximum output) It becomes important (in the part with a large inflow,
Since the generator output saturates, the error in the inflow prediction accuracy does not matter). In addition, the run-of-river power plant does not originally have a water storage function to prevent flooding in the downstream region, so it is not necessary to predict the flow rate to prevent flooding in the first place.

Since the conventional tank model and the storage function method cannot accurately predict from a portion with a large flow rate to a portion with a low flow rate, the prediction from the beginning of the increase in the flow rate to the peak (in a normal dam, flood It is designed so that the important part regarding safety such as prevention) is the best.
It is not impossible to adjust so that the prediction accuracy of the part with a small inflow is high, but these methods are
Individually designed according to each characteristic (parameter adjustment)
However, there is a problem that the development cost / introduction cost is high in any case.

On the other hand, in the method using the neural network, the prediction model can be constructed relatively easily because the neural network automatically learns only by giving input / output data. Also, its prediction accuracy is generally
Higher than tank model and storage function method.

However, as described above, in hydroelectric power generation in a self-propelled dam, since the generator output reaches maximum saturation long before the flood flow rate is reached, the conventional flow rate prediction model designed for flood prevention is Highly accurate forecast of power generation cannot be expected. This also applies to the flow rate prediction model using a neural network. Since the neural network learns so as to reduce the total learning error of all the learning data, learning including the data at the time of flood is the most important "range in which the amount of power generation changes" for predicting the power generation of a run-of-river power plant. Learning accuracy may be sacrificed (range where the generator output is maximally saturated; range where the inflow is relatively small compared to when a flood occurs).

If learning is performed using only the data in the "range in which the amount of power generation changes", highly accurate prediction of the amount of power generation in the "range in which the amount of power generation changes" can be expected. However,
Neural networks have the disadvantage that they do not know what kind of output will be given to data outside the learning range. Therefore, for example, there is a possibility that “the amount of power generation is small” may be predicted even during heavy rain.

After all, as for conventional self-propelled dams, the output of the neural network is used as the "inflow amount" as in the case of normal dams. There was a limit to improving the prediction accuracy of (the part where the inflow is relatively small).

An object of the present invention is to provide a neural network and a power generation amount prediction method capable of highly accurately predicting the power generation amount of a self-propelled dam.

[0016]

The invention according to claim 1 is
A method for predicting a hydroelectric power generation amount of a self-propelled dam using a neural network, wherein the output of the neural network is used as the hydroelectric power generation amount, and the hydroelectric power generation amount is directly predicted.

The invention according to claim 2 is a neural network used for predicting the hydroelectric power generation amount of a self-propelled dam, wherein the output is the hydroelectric power generation amount and the hydroelectric power generation amount is directly predicted. .

Unlike a normal dam, a power plant of a self-propelled dam does not originally have a water storage function for flood prevention in the downstream region, and therefore a flow rate prediction for flood prevention is unnecessary in the first place. Therefore, according to the inventions of claims 1 and 2, the power generation of the self-propelled dam can be performed by specializing for the self-propelled dam and directly predicting the power generation amount without temporarily predicting the flow rate. The quantity can be predicted with high accuracy. This effect has been confirmed by experiments.

Further, as the input data to the neural network, the rainfall amount in the upstream region, the river flow rate, the inflow amount to the dam, the temperature, the past power generation amount actual value, etc. may be used. When these difference data are used as in the described invention, it can be expected that the power generation amount can be predicted with higher accuracy.

Further, when the temperature is used as the input data, it becomes possible to make a more accurate prediction corresponding to the increase of the inflow amount by "snow melting water" especially in the spring season.

[0021]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. In order to improve the accuracy of power generation prediction in a self-propelled dam, the present invention specializes in predicting the power generation of a self-propelled dam, uses the output of a neural network as "power generation", and directly predicts the generator output. It is characterized by this. The input of the neural network is
It is the amount of rainfall in the upstream area, the amount of river flow, the amount of inflow to the dam, the temperature, the past actual amount of power generation, etc. However, it is possible to make more accurate predictions by using these difference data.

In this way, by setting the output of the neural network to the generator output (power generation amount), it is possible to make an appropriate prediction even when using data including past flood data as learning data.

First, the principle of the present invention that enables highly accurate prediction will be described. FIG. 1 shows an example of the correspondence relationship between the rainfall amount, the inflow amount, and the power generation amount. As described above, what is particularly important in predicting the hydroelectric power generation amount of a self-propelled dam is the “range in which the power generation amount changes”. In the example shown in FIG. 1, the rainfall amount 11 is between “5” and “25”. It is within the range (since the power generation amount 13 is '15', the maximum output (saturation) is reached). In predicting the power generation amount of a self-propelled dam, improving the prediction accuracy in this range is the most important.

In the example of FIG. 1, the inflow amount 12 is the rainfall amount 11
Always changes in the range of "0" to "40", and especially when the rainfall 11 becomes "25" or more, the amount of change in the inflow 12 increases. For normal dams (dams that are not run-of-river dams), this rainfall amount 1
It was important to predict the inflow rate 12 in the part where 1 is 25 or more. On the other hand, as for the power generation amount 13, the rainfall amount 11 is “5” to “2”.
Although it has changed in the 5'range, the rainfall 11 is '25'
Above, it saturates (the maximum output reaches '15').

When the neural network is trained with the rainfall amount shown in FIG. 1 as input data, in the conventional neural network in which the output is the "inflow amount", the inflow amount 12 is all "0" to "85". Learn the state.
Therefore, in order to reduce even the prediction error in the range that is not directly related to the power generation amount prediction, that is, in the range where the inflow amount 12 is '50' to '85', the inflow amount 12 is '1'.
The prediction error in the range of 0'to '50' does not always become small, and there is a limit.

On the other hand, it is possible to surely reduce the learning error in this range by learning only the data of the rainfall amount 11 in the range of '5' to '25', but the rainfall amount 11
However, there is a drawback in that it is not possible to guarantee what kind of output is performed for data other than "5" to "25" (outside the learning range).

When the output of the neural network is "power generation amount" as in this example, the rainfall amount 11
Is learning the data in the range of '0' to '40', but (in contrast to the "inflow amount"), the amount of rainfall 11 is in the range of '0' to '5', the rainfall amount. 11 is '25'
The power generation amount data corresponding to the range of “40” is in a saturated state (the values in the example in the figure are “0” 15 ”). Neural networks, for learning saturated regions,
Since learning error is difficult to appear in principle (the neural network uses a saturation function called sigmoid function,
Since it is used for the response function of the neuron, the learning of the saturated state is easy), and it is possible to concentrate the learning in the range of the rainfall amount 11 from '5' to '25' (the learning frequency is different). Even if the range is the same, it has a great effect on learning). Further, since the rainfall 11 has also learned for data outside the range of "5" to "25", the possibility of strange output for data outside the range is extremely low.

A simple experimental result demonstrating that the above principle is correct will be described below. In this experiment,
As shown in (1) and (2) below, the rainfall amount 11 shown in FIG.
Is the input data, and the output is the "inflow rate (predicted value)"
The learning result (prediction accuracy) of an actual experiment will be described below for two neural networks (conventional) and "power generation amount (predicted value)" (this example). (1) (Input, output) = (Amount of rainfall, amount of inflow) ... Conventional (2) (Input, output) = (Amount of rainfall, amount of power generation) ... Learning conditions for this example are (1), ( 2), as shown in FIG.
The number of times of learning is 10,000, and the number of units in the input layer / intermediate layer / output layer is one each. The neural network structure is not particularly shown because it is a simple example. As the teacher data, the data of the inflow amount 12 and the power generation amount 13 shown in FIG. 1 are used, respectively.

FIG. 3 shows learning results (prediction results) of the above two neural networks. Here, "a part where the amount of power generation changes" (a part important in predicting the amount of power generation of a self-propelled dam), that is, a range where the rainfall amount 11 is "5" to "25" is evaluated. This evaluation is made by averaging each of the “absolute value” 23 and the “absolute value” 26 (absolute values of “error” 22 and “error” 25) shown in FIG. However, in this example, when the rainfall amount 11 is "5", the error rate cannot be calculated because the power generation amount (actual value) 13 is "0", and therefore it is excluded from the evaluation targets.

As shown in FIG. 3, when the predicted value 21 of the inflow amount is used as the output of the neural network as in the conventional case, the average of the "absolute value" 23 of the error is 9.88 (%). . On the other hand, when the predicted value 24 of the power generation amount is used as the neural network output as in this example, the average of the “absolute value” 26 of the error is 4.95 (%).

As described above, the prediction error in the case of using the neural network according to this example is about half that in the conventional case, and the output of the neural network is set to "the amount of power generation". It has been confirmed to show good results.

The error is (predicted value-actual value) /
Calculate based on actual values. Although the predicted inflow amount and the predicted power generation amount are compared in FIG. 3, as described above, conventionally, the predicted inflow amount is applied to the QP characteristics of the hydropower generator to predict the power generation amount. Therefore, of course, the error of the predicted inflow is directly linked to the error of the predicted power generation.

As described above, according to this example, the prediction accuracy could be greatly improved by simply setting the output of the neural network to "the amount of power generation". In addition, the conventional method for predicting the amount of hydroelectric power generation is to first predict the inflow to the dam,
Next, this predicted inflow was converted into predicted power generation. That is, two-step processing was required. In the neural network of this example, since the power generation amount is directly predicted, the processing is simplified and the processing time can be shortened.

Below, while showing some examples of the neural network structure according to this example, the actual experimental results show that
It is confirmed that the prediction accuracy is improved.
First, with reference to FIGS. 4 and 5, an example (part 1) in which the power generation amount is actually predicted using the neural network according to the present example will be described.

First, FIG. 4A shows input data / output data of the neural network according to this example. As shown in the figure, as input data, difference data of rainfall amounts for the past several hours and difference data of generator output (actual value) are used. The output data is the generator output (predicted value) difference from 1 hour ahead to 3 hours ahead. Note that "generator output" has the same meaning as "power generation amount".

In this example, difference data (rainfall difference, generator output (actual value) difference) is used as input data, but the present invention is not limited to this. The rainfall amount and the generator output (actual value) for the past several hours may be used. In this case, naturally, the output data is the generator output (predicted value) from 1 hour to 3 hours. However, it can be expected that the use of the difference data will make it possible to predict the power generation amount with higher accuracy. The method of using difference data as input data has already been disclosed by the present applicant, for example, in Japanese Patent Laid-Open No.
0-260718 Publication "Water amount prediction method in dam"
Have already been proposed.

Learning conditions are shown in FIG. 4 (b). In this example, the number of times of learning is 100,000, the number of units in the input layer is 22, the number of units in the intermediate layer is 1, and the number of units in the output layer is 3.

As the learning method, in this example, the learning method described in Japanese Patent Application No. 2000-71001 “Neural network optimization learning method” previously proposed by the applicant of the present application is adopted. , But not limited to this. still,
This learning method has a function of reducing the number of elements in the intermediate layer. Therefore, the number of elements in the intermediate layer is reduced during learning, and at the end of learning, as shown in FIG. The number is one, but this is only an example.

FIG. 4C shows an example of the neural network under the above conditions. In this example, the input layer 2
Of the two units, 19 units are for inputting rainfall difference data, and 3 units are for inputting generator output (actual value) difference data. Of course, this is just one example. As the rainfall amount difference data, a difference value with the rainfall amount at each time point in the past 1 hour (1 h), 2 h ago, ... 19 h ago is used based on the rainfall amount at the current time. The same applies to the generator output (actual value) difference data, which is based on the generator output (actual value) at the current time, one hour before (1h) in the past, and 2h.
The difference value with the generator output (actual value) at each time before and 3 hours before is used. Each of the three units in the output layer has a generator output (predicted value) one hour ahead (1h) / 2h ahead / 3h ahead
Output the difference value. Also in this case, the difference value between the generator output (actual value) at the current time and the generator output (predicted value) 1 hour (1h) ahead / 2h ahead / 3h ahead is output.

The absolute value of the difference between the value obtained by outputting the generator output (predicted value) difference using the above neural network and the actual generator output 1h ahead / 2h ahead / 3h ahead is calculated, and the average value thereof is calculated. The result obtained is shown in FIG. As shown in the figure, the average value of the absolute values of the errors (herein referred to as the absolute value average error) gives good results from 1 hour ahead prediction to 3 hours ahead prediction, and particularly 1 hour ahead prediction %, Which is a very good result.

FIG. 5 (b) is a graph showing the result of continuing to predict the amount of power generation one hour ahead of time for a long period of about 781 hours, and the actual amount of power generation after one hour. I will show you.

As shown in the figure, the two are almost overlapped with each other, and the result is so good that the error cannot be discriminated by looking at this. Next, with reference to FIGS. 6 and 7, an example (part 2) in which the power generation amount is actually predicted using the neural network according to the present example will be described.

In this example, as the input data, in addition to the difference data of the rainfall amount for the past several hours and the difference data of the generator output (actual value), the difference data of the temperature is used. Here, the meaning of adopting the temperature as an input factor will be described.

The temperature used as input data is
This is especially effective when making predictions in "spring." That is, when snow is accumulated in the upstream region of the self-propelled dam, the inflow amount tends to increase due to so-called “melting water” as the temperature rises. Therefore, by using the temperature as one of the input factors, we were able to respond to the increase in inflow by "snow melting water".
It is possible to make a more accurate prediction.

In addition to the above-mentioned data, the amount of rainfall in the upstream region, the flow rate of the river, etc. (and the difference data thereof) may be used as the input data. FIG. 6 shows an example of the neural network structure according to this example.

In the illustrated example, the number of units in the input layer is 22.
1h before, 2h before ...
Rainfall difference data before 10h, before 1h, before 2h, ...
., 6h before generator output (actual value) difference data, 1h
The temperature difference data before 2h, ..., 6h before is input. Similar to the above, the number of units in the intermediate layer is one as a result of the reduction in the number of units during learning. In this example, the number of units in the output layer is one, and the generator output (predicted value) difference value one hour ahead is output.

As a result of actual prediction using this neural network, as shown in FIG. 7A, the one-hour ahead prediction has an absolute value average error of 0.85%, which is very good. Good results have been obtained.

FIG. 7 (b) shows the relationship between the result of the prediction of the amount of power generation one hour ahead of time during the period of about 313 hours, the actual amount of power generation after one hour, and the temperature. Is shown as a graph. In the illustrated example, the power generation amount slightly increases near 60 hours and around 280 hours. By referring to the temperatures around 60 hours and around 280 hours, it can be confirmed that the high temperature continues for a relatively long time, and it can be seen that the temperature slightly affects the power generation amount. However, the same figure shows an example of forecasting in December. As mentioned above, "spring" has the greatest effect of temperature. It is certain that the impact on quantity will be greater.

FIG. 8 is a hardware configuration diagram of an information processing apparatus which is an example of an implementation mode of a method for predicting the power generation amount of a self-propelled dam using the above-mentioned neural network.
It should be noted that this information processing apparatus is installed, for example, in a system control station, a power supply command station, a dam management station, a hydroelectric power station, or the like. Also,
This information processing apparatus is, for example, a network connection device 36, which will be described later, through some communication line, a rain gauge installed in the upstream region, a flow meter for measuring the flow rate of the upstream river,
It is configured so that data can be acquired from a flow meter that measures the flow rate into the dam, a thermometer that measures the temperature, and a generator.

The information processing apparatus 30 shown in FIG.
1, a storage device 32, an input device 33, an output device 34, a medium drive device 35, a network connection device 36, etc., and these are connected to a bus 37. The configuration shown in the figure is an example, and the present invention is not limited to this.

The CPU 31 is a central processing unit that controls the entire information processing apparatus 30, and a storage device 32 described later.
Various processes are executed based on the data / program stored in the etc.

The storage device 32 is, for example, an HDD or a ROM.
/ RAM / memory such as flash memory, which stores a database structure of the neural network (number of input layer / intermediate layer / output layer elements, weight size, input data, output data, etc.). It also stores a program that causes the CPU 31 to execute various processes.

The input device 33 is, for example, a keyboard, a mouse or the like. The output device 34 is a display or the like, and displays the calculation result of the neural network described above and the like.

The medium driving device 35 reads out programs / data and the like stored in the portable storage medium 38. The portable storage medium 38 is, for example, an FD (floppy disk).
38a, CD-ROM 38b, DVD, magneto-optical disk and the like. The program / data stored in the storage device 32 may be stored in the portable storage medium 38. That is, the various processes described above are
The programs / data and the like stored in the portable storage medium 38 are processed by the information processing device 30 via the medium driving device 35.
It may be loaded on the side and executed.

Further, the network connection device 36 may be configured to connect to a network (not shown) such as the Internet to enable transmission / reception of programs / data etc. to / from an external information processing device etc. The program / data may be one that downloads the program / data stored in the storage device of the device on the external information provider side via the network connected by the network connection device 36.

The present invention can be configured as a storage medium (portable storage medium 38 or the like) storing the above program / data itself, or as a program itself.

[0057]

As described above in detail, according to the neural network of the present invention and the prediction method using the neural network, the neural network output is made to be "power generation amount" by specializing for the self-current dam. , High prediction accuracy and low development cost.

[Brief description of drawings]

FIG. 1 shows an example of the correspondence between rainfall, inflow, and power generation.

FIG. 2 is a diagram showing learning conditions in a simple example.

FIG. 3 is a diagram showing comparison of learning results by a neural network of a conventional example and a neural network of the present example.

4A to 4C are diagrams for explaining an example (part 1) in which the power generation amount is actually predicted using the neural network according to the present example.

5A and 5B are diagrams showing prediction results.

FIG. 6 is a diagram for explaining an example (part 2) in which the power generation amount is actually predicted using the neural network according to the present example.

7A and 7B are diagrams showing prediction results.

FIG. 8 is a hardware configuration diagram of an information processing apparatus, which is one embodiment of a method for predicting the power generation amount of a self-propelled dam using the neural network of the present example.

[Explanation of symbols]

11 rainfall 12 Inflow 13 Power generation 21 Inflow (predicted value) 22 Error 23 Absolute value (of error) 24 Electricity generation (predicted value) 25 error 26 Absolute value (of error) 30 Information processing equipment 31 CPU 32 storage 33 Input device 34 Output device 35 medium drive 36 Network connection device 37 bus 38 Portable storage medium 38a FD (floppy disk) 38b CD-ROM

Claims (3)

[Claims] 1. A method for predicting hydroelectric power generation of a self-propelled dam using a neural network, wherein the output of the neural network is used as the hydroelectric power generation amount, and the hydroelectric power generation amount is directly predicted. Method for predicting hydroelectric power generation of dams. 2. A neural network used for predicting the hydroelectric power generation amount of a self-propelled dam, characterized in that the output is the hydroelectric power generation amount and the hydroelectric power generation amount is directly predicted. network. 3. The input data to the neural network includes rainfall in the upstream area, river flow rate, inflow rate to the dam,
The neural network according to claim 2, wherein the differential data is at least one of the temperature and the past power generation amount actual value.