JP2005100972A - Fuel cell system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a fuel cell system improving an energy saving property by responding to power loads which are different with homes. <P>SOLUTION: A power load accumulating part 8 accumulates histories of power load track records. A control method determining part 13 selects a control method for amount of power generation based on items characterizing power load patterns different with homes or seasons or power load pattern for control time bands sectioned off according to the power load pattern from the power load histories accumulated by the power load accumulating part 8. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a fuel cell system that supplies electric power using fuel gas.

  Conventionally, a fuel cell system that supplies electric power and hot water using fuel gas is known. In this fuel cell system, since the power generation efficiency increases when the power generation output is high, it is not desirable that the power generation amount is significantly lower than the power consumption and the operation is performed at a low output, and the utilization rate of the fuel cell system also decreases. As a result, the energy saving performance is impaired. In addition, if the amount of power generation exceeds the amount of power used, it is possible to sell the surplus power to the power company. However, since the unit price of power sales is generally low, it is desirable that the power generation cost will deteriorate as a result. Absent.

  Therefore, it is conceivable to start / stop the system in accordance with a predetermined household power usage pattern. In power plants and the like, a technique for predicting an electric power load from information such as climate, temperature, day of the week, etc. has been devised from early on (see, for example, Patent Document 1 and Patent Document 2). However, the power load pattern in the home varies greatly from home to home, and there are various fluctuating factors. Predetermining the parameters that determine the load pattern, such as climate, temperature, and day of the week, is used in the home. This is not sufficient for a fuel cell system that assumes that In order to solve such a problem, the conventional fuel cell system performs power generation control following the load of power used. That is, the fuel cell system is activated at a ratio of time exceeding a certain power load within a certain time that has elapsed, and after the activation, the power generation amount is increased or decreased following the power load. The stop timing is determined by the proportion of time that falls below a certain power load within a certain period of time as in the case of activation.

  In addition, in the fuel cell system, power generation is performed according to the load of electric power as one of the operation controls for increasing the power generation efficiency during operation. However, when the power generation output is small, there is a problem that the power generation efficiency is poor. In some cases, it is better in terms of energy and cost to purchase power from an electric power company than to generate power with the system.

For this reason, there is a DSS (Daily Start and Stop) that stops the system in a time zone with a low power load and starts it when the power load exceeds a certain amount. In the fuel cell system using the DSS, control for operating, starting and stopping the system is performed following the load of the electric power used. In determining whether to start the system, the proportion of time exceeding a certain power load is used within a certain period of time, and the amount of power generation is increased or decreased following the power load after startup. For the determination of the stop, as in the case of the start, the ratio of the time that falls below a certain power load within a certain time that has elapsed is used. However, in this method, since a certain time is required for the start / stop determination, a time lag occurs with respect to the generation and reduction of the power load. For this reason, recently, a method has been proposed in which a load is predicted by a prediction method such as case-based reasoning using information such as past power load, hot water supply load, temperature, and calendar, and operation control is judged based on the predicted result. (For example, see Patent Document 3). The power load in a home varies from home to home, and the pattern of occurrence varies from day to day in a single home. Therefore, the prediction method using case-based reasoning that makes use of past cases that occurred in the home is an effective method. it is conceivable that.
Japanese Patent Laid-Open No. 3-164804 JP-A-9-273895 JP 2003-329292 A

  However, in a fuel cell, power cannot be generated unless hydrogen gas is generated after the reformer for generating hydrogen-based gas by reforming fuel such as city gas is heated to a high temperature. It takes about one hour from the start of the fuel cell system to the actual power generation. Therefore, if the system is started after a load is generated according to the conventional control method, it is difficult to cover the power load generated in the time elapsed until the start and the power load generated from the start to the generation of power. It is. In addition, in the case of a stop, it is possible to stop the system immediately, but the conventional method requires a determination time to stop even when the power load is reduced. This may impair the energy saving performance of the fuel cell system.

  In addition, case-based reasoning requires that a large amount of memory capacity must be secured to accumulate information in order to infer future data from past large amounts of power demand, hot water supply demand, room temperature, and calendar data. There is a problem that it takes a calculation cost such as a calculation time to perform. In fact, for households where the load generation pattern is almost fixed, for example, a simple prediction method such as arithmetic averaging can be used for prediction with sufficient accuracy. Complex processing such as base inference is not always necessary.

  The present invention has been made to solve the above problems, and an object of the present invention is to provide a fuel cell system that can cope with different power loads in each home and can improve energy saving. is there.

  In order to solve the above-described conventional problems, a fuel cell system according to the present invention is a fuel cell system that supplies power using fuel gas, the power load accumulation means for accumulating a history of power load results, Control method determination means for determining a control method for controlling the amount of power generation from the power load pattern based on the power load history accumulated by the power load accumulation means, and power generation control is performed according to the control method determined by the control method determination means Power generation control means.

  According to this configuration, a history of power load results is accumulated, a method for controlling the amount of power generation is determined from a power load pattern based on the accumulated power load history, and power generation control is performed according to the determined control method.

  The fuel cell system further includes a variation coefficient calculating unit that acquires a power load history accumulated by the power load accumulation unit and calculates a variation coefficient representing a variation degree of the power load pattern, and determines the control method. The means preferably determines a control method based on the variation coefficient calculated by the variation coefficient calculation means.

  According to this configuration, the power load history is acquired, the variation coefficient indicating the degree of variation of the power load pattern is calculated, and the control method is determined based on the calculated variation coefficient.

  Further, in the above fuel cell system, the variation coefficient calculating means acquires the power load history accumulated by the power load accumulating means, the rising time when the power load increases, the rising slope when the power load increases, the power load A variation coefficient indicating a variation amount of each of the electric energy at the time when the peak appears and the peak of the power load is calculated, and the control method determining means is one of the variation coefficients calculated by the variation coefficient calculating means. It is preferable to determine the control method from a plurality of combinations.

  According to this configuration, the power load history is acquired, the rising time when the power load increases, the rising slope when the power load increases, the time when the peak of the power load appears, and the degree of variation of the power amount at the peak of the power load. A variation coefficient to be shown is calculated, and a control method is determined from one or a combination of the calculated variation coefficients.

  Further, in the fuel cell system, the control method determining means has a variation coefficient of rise time at which the power load increases that is a predetermined value or less, and a variation coefficient of the electric energy at the peak of the power load is a predetermined value or less. In this case, it is preferable to determine the control method using the arithmetic average of the power load amount for each time from the power load history.

  According to this configuration, when the variation coefficient of the rise time at which the power load increases is equal to or smaller than the predetermined value and the variation coefficient of the power amount at the peak of the power load is equal to or smaller than the predetermined value, the power for each time is calculated from the power load history. The control method using the arithmetic average of the load is determined.

  Further, in the above fuel cell system, the control method determining means has a rise time variation coefficient for increasing the power load equal to or greater than a predetermined value, and a rise gradient variation coefficient for increasing the power load is equal to or greater than a predetermined value. When the variation coefficient of the time at which the peak of the power load appears is a predetermined value or more and the variation coefficient of the power amount at the peak of the power load is a predetermined value or more, the past power load pattern is followed from the power load history. It is preferable to determine the control method.

  According to this configuration, the variation coefficient of the rise time at which the power load increases is equal to or greater than the predetermined value, the variation coefficient of the rise gradient at which the power load increases is equal to or greater than the predetermined value, and the peak of the power load appears. When the variation coefficient is equal to or greater than a predetermined value and the variation coefficient of the power amount at the peak of the power load is equal to or greater than the predetermined value, the control method is determined to follow the past power load pattern from the power load history.

  Further, in the above fuel cell system, when the conditions described in claims 4 and 5 are not satisfied, it is preferable to determine a control method using a neural network.

  According to this configuration, the condition that the variation coefficient of the rise time at which the power load increases is equal to or smaller than the predetermined value and the variation coefficient of the power amount at the peak of the power load is equal to or smaller than the predetermined value, and the power load increases. The rise time variation coefficient is greater than or equal to a predetermined value, the rise gradient variation coefficient at which the power load increases is greater than or equal to the predetermined value, and the time variation coefficient at which the peak of the power load appears is greater than or equal to the predetermined value and power When the condition that the variation coefficient of the electric energy at the peak of the load is not less than a predetermined value is not met, the control method is determined using a neural network.

  Further, in the above fuel cell system, when the method for controlling the power generation amount determined by the control method determination unit is a control method using a neural network, based on the variation coefficient calculated by the variation coefficient calculation unit, It is preferable to further include a neuro model selection unit that selects a neuro model of the neural network. According to this configuration, when the method for controlling the power generation amount is a control method using a neural network, the neural network neuro model is selected based on the variation coefficient.

  In the above fuel cell system, it is preferable that the neuro model selection unit selects a combination of an input parameter and an output parameter for constructing the neuro model according to the neuro model. According to this configuration, a combination of input parameters and output parameters for constructing a neuro model is selected according to the neuro model.

  Further, in the fuel cell system, the variation coefficient calculating unit calculates a variation coefficient indicating a variation degree of the power load amount at the same time in a certain period, and the control method determining unit is configured by the variation coefficient calculating unit. When the calculated variation coefficient is less than a predetermined value, the control method using the arithmetic average of the power load amount at each time is determined from the power load history, and the variation coefficient calculated by the variation coefficient calculation means is greater than or equal to the predetermined value The power generation control unit calculates a predicted load amount using the control method determined by the control method determination unit, and performs power generation control according to the calculated predicted load amount. It is preferable.

  According to this configuration, the variation coefficient indicating the variation degree of the power load amount at the same time in a certain period is calculated, and when the calculated variation coefficient is less than a predetermined value, the power load amount for each time is calculated from the power load history. When the calculated variation coefficient is determined to be a control method using an arithmetic average and the calculated variation coefficient is equal to or greater than a predetermined value, the control method is determined using a neural network, and the predicted load amount is calculated and calculated using the determined control method. Power generation control is performed according to the predicted load amount.

  Further, in the fuel cell system, the power load history accumulating unit acquires the outside air temperature from an outside air temperature sensor that measures the outside air temperature, accumulates the acquired outside air temperature together with the power load amount, and the power generation control unit includes: When the control method determining means determines a control method using a neural network, the predicted load using a neuro model with the power load amount and the outside air temperature as input values and the predicted load amount of the power generation amount as an output value It is preferable to calculate the amount and perform power generation control according to the calculated predicted load amount.

  According to this configuration, the outside air temperature is acquired from the outside air temperature sensor that measures the outside air temperature, and the acquired outside air temperature is accumulated together with the power load amount. When the control method using the neural network is determined, the predicted load amount is calculated using a neuro model in which the power load amount and the outside air temperature are input values, and the predicted load amount of the power generation amount is an output value. Power generation control is performed according to the predicted load amount.

  Further, in the fuel cell system, the power load history accumulated by the power load history accumulating unit is acquired, a day is divided into a plurality of divided time zones, and a start time and an end time for each divided time zone. A time zone dividing unit that generates divided time zone information that is a combination of the power load history, the variation coefficient calculating unit, the power load history accumulated by the power load accumulation unit, and the time zone dividing unit. The division time zone information is obtained, a variation coefficient representing a variation degree of the power load amount is calculated for each division time zone, and the control method determination means is a variation coefficient calculated by the variation coefficient calculation means. And the divided time zone information generated by the predicted time division means, and for each divided time zone based on the variation coefficient, Determining a control method, and the power generation control means calculates a predicted load amount for each of the divided time zones using the control method determined by the control method determination means, and performs power generation control according to the calculated predicted load amount. Preferably it is done.

  According to this configuration, the accumulated power load history is acquired, and divided time zone information that is a combination of the start time and end time for each divided time zone obtained by dividing one day into a plurality of times is generated. The accumulated power load history and the generated divided time zone information are acquired, and a variation coefficient representing the degree of variation in the power load amount is calculated for each divided time zone. The calculated variation coefficient and the generated divided time zone information are acquired, the control method is determined for each divided time zone based on the variation coefficient, and the predicted load for each divided time zone using the determined control method The amount is calculated, and power generation control is performed according to the calculated predicted load amount.

  Further, in the above fuel cell system, the time zone dividing unit averages the power load amount for a certain period from the power load history acquired from the power load accumulation unit at the same time, and is averaged. It is preferable to divide the time from the time when the power load amount at each time changes from the fall to the rise to the time when the power load changes from the next fall to the rise as one division time zone.

  According to this configuration, in the acquired power load history, the power load amount for a certain period is averaged at the same time, and the averaged power load amount at each time is changed from the falling time to the rising time. The time from the next falling to the rising time is divided as one divided time zone.

  Further, in the above fuel cell system, the control method determination means selects either one of start and stop according to the ratio of the time exceeding the reference power amount in the divided time zone from the power load history. Preferably, it is determined.

  According to this configuration, the control method is selected to select either start or stop according to the ratio of the time exceeding the reference power amount in the divided time zone from the power load history.

  In the fuel cell system described above, it is preferable that the variation coefficient calculating means calculates the variation coefficient of the power load amount at the same time on a plurality of days when determining the variation coefficient. According to this configuration, when obtaining the variation coefficient, the variation coefficient of the power load amount at the same time on a plurality of days is obtained by averaging.

  Further, in the fuel cell system, the power load history accumulated by the power load history accumulating unit is acquired, and the acquired power load history is an exceptional day that is an exceptional load pattern different from other days. It is preferable to further include an exception date determination unit that controls whether or not to store the power load amount of the exception day in the power load history stored by the power load storage unit when it is determined whether there is an exception date. .

  According to this configuration, the power load history is acquired, and it is determined whether or not the acquired power load history is an exceptional day that is an exceptional load pattern different from other days. Control is performed so that the power load amount on the exceptional day is not accumulated in the power load history.

  According to the fuel cell system of the present invention, it is possible to cope with different power load patterns in individual households, and energy saving can be improved.

  Embodiments according to the present invention will be described below with reference to the drawings. In addition, about the same structure in each figure, the same code | symbol is attached | subjected and the description is abbreviate | omitted.

(First embodiment)
FIG. 1 is a diagram showing an overall configuration of a fuel cell system according to the present invention. The fuel cell system shown in FIG. 1 includes a fuel cell 1, a power output line 2, a load 3, a commercial power source 4, a power meter 5, and a control unit 6.

  The fuel cell 1 converts chemical energy into electrical energy by reacting hydrogen obtained from fuel such as city gas with oxygen in the air. Various household electrical appliances are connected as a load 3 to the power output line 2 from the household fuel cell 1. A commercial power source 4 is connected to the power output line 2 as a power system. The power output line 2 is provided with a wattmeter 5 that measures the power output from the fuel cell 1 and the commercial power supply 4. The wattmeter 5 is connected to the control unit 6, and the control unit 6 is connected to the fuel cell 1 that receives the power generation amount at the next time.

  FIG. 2 is a block diagram illustrating a configuration of the control unit 6 in the first embodiment. 2 includes a power load accumulation unit 8, an exception date determination unit 9, a power load history storage unit 10, a control time zone division unit 11, a variation coefficient calculation unit 12, a control method determination unit 13, and a constant control unit. 14, a load prediction unit 151, an arithmetic mean control unit 152, a follow-up control unit 16, a neuro model selection unit 17, a load prediction unit 181, a neuro control unit 182, and a neuro model storage unit 19.

  The power load accumulation unit 8 accumulates the power load history output from the power meter 5 in the power load history storage unit 10.

  The exception date determination unit 9 determines whether or not it is an exceptional day that is an exceptional load pattern. If it is an exception date, the power load history accumulated in the power load history storage unit 10 by the power load accumulation unit 8 In such a case, control is performed so that the power load record on the exceptional day is not accumulated.

  The power load history storage unit 10 stores a history of power load results. The control time zone dividing unit 11 divides one day, that is, 24 hours into a plurality of control time zones based on the average power load of the accumulated power for one hour.

  The variation coefficient calculation unit 12 includes a rise time, a rising gradient, a time when a power load peak appears, and a power load peak when the power load suddenly increases from the power load history accumulated in the power load history storage unit 10 by the power load accumulation unit 8. A coefficient of variation indicating the degree of variation of each power amount at the peak is calculated. In the present embodiment, the variation coefficient indicating the degree of variation is calculated using a variation coefficient. However, the present invention is not particularly limited thereto, and the variation coefficient may be calculated using another coefficient.

  The control method determination unit 13 determines a method for controlling the power generation amount based on items characterizing the power load pattern from the power load history accumulated in the power load history storage unit 10 by the power load accumulation unit 8. That is, the control method determination unit 13 is based on the variation coefficient calculated by the variation coefficient calculation unit 12, the constant control unit 14, the arithmetic average control unit 152 (load prediction unit 151), the follow-up control unit 16, and the neuro control unit 182. A method for controlling the power generation amount is determined by calling one of the (load prediction unit 181).

  The constant control unit 14 controls whether to continue the power generation or to stop the power generation over the entire control time zone for controlling the fuel cell 1.

  The load prediction unit 151 obtains the predicted load power at the next time by obtaining the average of each time of the power load history stored in the power load history storage unit 10. The arithmetic average control unit 152 transmits the predicted load power amount derived by the load prediction unit 151 to the fuel cell 1 as the required power generation amount.

  The follow-up control unit 16 refers to the power load history immediately before being stored in the power load history storage unit 10 and starts the fuel cell system at a ratio of time exceeding a certain power load within a certain time. Follows the power load to increase or decrease the amount of power generation, and the stop timing is determined by the ratio of the time that falls below a certain power load within a certain time that has passed, as in the case of activation.

  The neuro model selection unit 17 selects a neuro model of the neural network stored in the neuro model storage unit 19 based on the variation coefficient calculated by the variation coefficient calculation unit 12.

  The load prediction unit 181 obtains a predicted power amount based on the neuro model selected by the neuro model selection unit 17. The neuro control unit 182 transmits the value to the fuel cell 1 using the predicted power amount obtained by the load prediction unit 181 as the required power generation amount at the next time.

  The neuro model storage unit 19 stores a plurality of neuro models prepared in advance. The neuro model stored in the neuro model storage unit 19 includes a time prediction model and a fluctuation prediction model. Each model will be described later.

  Next, the operation of the fuel cell system according to the present invention will be described.

  First, in the load learning state during the initial operation of the system, the power load storage unit 8 stores the amount of power measured by the power meter 5 in the power load history storage unit 10. In the present embodiment, the period of data used for exception date determination and load pattern learning is, for example, four weeks (28 days), and power generation control during that period is performed by the tracking control unit 16. After the end of the load learning state, processing described later is sequentially performed using load data for the latest four weeks among the data stored by the power load storage unit 8. In the present embodiment, the usage data period is set to 4 weeks in consideration of the variation of the pattern depending on the season, but the present invention is not limited to this.

And if it enters into a full operation period after passing through a predetermined load learning period, the control time slot | zone division | segmentation part 11 will read the load pattern of the latest 4 weeks from the electric power load log | history memory | storage part 10, and will be 1 day according to the read load pattern ( 24 hours) is divided into a plurality of control time zones. FIG. 3 is a diagram for explaining the control time period divided into a plurality of parts. In FIG. 3, the vertical axis represents the amount of power, and the horizontal axis represents time. As shown in FIG. 3, the control time zone dividing unit 11 obtains an average power load of integrated power every hour, and the time from the time when the average power load becomes a downward peak to the time when the next downward peak becomes 1 As a control time zone, 24 hours are divided into a plurality of control time zones. In other words, the control time zone dividing unit 11 examines the time when the slope of the characteristic indicating the average power load of the integrated power every hour changes from the downward slope to the upward slope, and 24 hours at each time. Split. In FIG. 3, first, the average power load has a downward peak at around 5 o'clock, the next downward peak at around 12:00, and the further downward peak at around 15:00. Therefore, the control time zone dividing unit 11 sets the first control time zone T ₁ from 5 o'clock to 12 o'clock, and the second control time zone T ₂ from 12 o'clock to 15 o'clock, from 15 o'clock to 5 o'clock. the split on the third three control time slot for 24 hours as a control time slot T ₃ of. In FIG. 3, 24 hours are divided into three at three times of 5 o'clock, 12 o'clock and 15 o'clock, but this time and the number of divisions are not limited depending on the load pattern.

  Next, the variation coefficient calculation unit 12 includes a time at which the power amount for each control time period divided into a plurality by the control time period division unit 11 reaches a peak, a load power amount at the peak time, and a load rising time. Then, a variation coefficient, a sample average, and a sample deviation indicating variations between the gradient of the rise time of the load and the numerical value of the electric energy that does not exceed a preset reference value are obtained. The variation coefficient is generally obtained by the following equation (1).

Cv (t) = (s / μ) × 100 (1)
In the above equation (1), Cv (t) represents the coefficient of variation of the past power load amount at time t (0 to 23:00), s represents the standard deviation of the past power load amount at time t, μ represents the average power load at the past time t. Further, the average μ is expressed by the following formula (2), and the dispersion s ² is expressed by the following formula (3).

μ = ΣXi / n (2)
s ² = Σ (Xi−μ) ² / n (3)
In the above formulas (2) and (3), Xi represents the power load amount on day i at time t corresponding to the number of accumulation days, and n represents the number of accumulation days (28 days).

Subsequently, control time slot for operation of the fuel cell system becomes a first control time slot T ₁ is the first day, the more power the load power meter 5 are transmitted to the power load storage unit 8, The exception date determination unit 9 determines whether or not the exception date is an exceptional load pattern.

FIG. 4 is a diagram for explaining the exception date determination in the present embodiment. As shown in FIG. 4, each of the load rise time and the gradient of the load rise time obtained by the variation coefficient calculation unit 12 has a normal distribution. The exception date determination unit 9 determines that an exception occurs when both the rise probability of the load rise time and the gradient of the rise time of the load obtained by the variation coefficient calculation unit 12 are 30% or less (regions R1 and R2 shown in FIG. 4). The power load information for the day is discarded, and otherwise, the load power amount is stored in the power load history storage unit 10. The next day, the exception date determination unit 9 makes a determination similar to the first control time slot T ₁ again. In the distribution example shown in FIG. 4, the factor value portion corresponds to the load rise time or the gradient value of the load rise time, and each value is applied to the regions R1 and R2 whose appearance probability is 30% or less. Does not accumulate load power.

  Then, the control method determination unit 13 calculates the ratio of time exceeding a preset reference value and the variation coefficient calculation unit 12 for each control time zone divided into a plurality of times by the control time zone division unit 11. The control method is determined according to the coefficient of variation. Note that the relationship between the value of the coefficient of variation, each control method, and the neuro model is prepared in advance as a combination that increases the power amount prediction accuracy based on the result of the power measurement experiment.

  First, the control method determination unit 13 includes a constant control unit 14 that controls whether to continue power generation or to stop power generation over all control time periods divided into a plurality of times by the control time period division unit 11. Make a call decision. When there is a power load of a certain amount or more over all the control time periods divided by the control time period dividing unit 11, it is not necessary to make a start / stop determination within each control time period, and always generate power. You can continue. In addition, there is a problem that the power generation efficiency cannot be increased unless the power generation continues beyond a certain amount due to the performance characteristics of the fuel cell. Therefore, in the control time zone, the control method determination unit 13 has a ratio of time when the power load exceeds the reference value of 400 kWh is 80% or more, or a ratio of time when the power load falls below 400 kWh is 80% or more. In this case, the method for controlling the power generation amount is determined to be constant control, and the constant control unit 14 is called. The determination reference value of 400 kWh is a value calculated from the power generation efficiency so that the power generation cost does not exceed the commercial power supply purchase price by comparing the economics of the fuel gas charge and the commercial power supply charge. When the ratio of the time exceeding the reference value is 80% or more, the constant control unit 14 transmits information on starting or continuing power generation to the fuel cell 1. Further, the constant control unit 14 transmits information on the stop of power generation to the fuel cell 1 when the ratio of the time below the reference value is 80% or more.

  The control method determination unit 13 calculates an arithmetic average as a method of controlling the power generation amount when the value of the coefficient of variation of the power rise time is equal to or less than a predetermined value and the value of the coefficient of variation of the peak power amount is equal to or less than the predetermined value The control is determined and the arithmetic average control unit 152 is called. Specifically, the control method determination unit 13 calls the arithmetic average control unit 152 when the variation coefficient of the power rise time is, for example, 8 or less and the variation coefficient of the peak power amount is, for example, 5 or less. The load prediction unit 151 calculates the average of the power load at each time from the power load history stored in the power load history storage unit 10 and obtains the predicted load power at the next time. The arithmetic average control unit 152 transmits the predicted load power amount obtained by the load prediction unit 151 to the fuel cell 1 as the required power generation amount. When the predicted load power amount exceeds the reference power amount at the time of power generation, the fuel cell 1 prepares for starting. When the predicted load power amount is lower than the reference power amount at the time of stop, the fuel cell 1 sets the stop timing. Can be measured. The power usage pattern in the case of using the control method by the arithmetic average control unit 152 is a case where almost the same life is taken every day, and the amount of power used can be predicted without requiring a complicated method. is there.

  As described above, when the variation coefficient of the rise time at which the power load increases is equal to or smaller than the predetermined value and the variation coefficient of the power amount at the peak of the power load is equal to or smaller than the predetermined value, the power load amount at each time is determined from the power load history. Therefore, when there is little variation in the power load pattern, the calculation cost for predicting the power load can be kept relatively small.

  The control method determination unit 13 determines that the value of the coefficient of variation of the power rise time is equal to or greater than a predetermined value, the value of the coefficient of variation of the power peak time is equal to or greater than a predetermined value, and the coefficient of variation of the peak power amount. When the value is equal to or greater than the predetermined value and the value of the variation coefficient of the rising slope is equal to or greater than the predetermined value, the method for controlling the power generation amount is determined as follow-up control, and the follow-up control unit 16 is called. Specifically, the control method determining unit 13 determines that the value of the coefficient of variation of the power rise time is, for example, 16 or more, the value of the coefficient of variation of the power peak time is, for example, 10 or more, and the peak power amount When the value of the fluctuation coefficient is 30 or more and the fluctuation coefficient value of the rising gradient is 60 or more, for example, the follow-up control unit 16 is called. The control method performed by the follow-up control unit 16 is a conventional control method, and does not require a past power history, and starts and stops power generation only with the previous power history. Since this control method requires a judgment time of about one hour immediately before, there is a possibility that it cannot sufficiently respond to the start / stop of power generation, but it lives in a different pattern every day, and the power usage pattern varies widely. When it is too impossible to predict, energy saving can be improved by controlling by this method.

  When the variation coefficient value calculated by the variation coefficient calculation unit 12 does not apply to the above two control methods (arithmetic average control and follow-up control), the control method determination unit 13 sets the method for controlling the power generation amount to neuro control. Then, the neuro model selection unit 17 is called. The neuro model selection unit 17 selects a combination of a plurality of neuro model input parameters and output parameters prepared in advance in the neuro model storage unit 19 in accordance with the value of the variation coefficient calculated by the variation coefficient calculation unit 12. In the present embodiment, the neuro model storage unit 19 stores in advance two neuro models, a time prediction model and a fluctuation prediction model.

  FIG. 5 is a diagram for explaining the input / output parameters of the neuro model in the present embodiment. FIG. 5A is a diagram for explaining the input / output parameters of the time prediction model in the present embodiment. FIG. 5B is a diagram for explaining input / output parameters of the fluctuation prediction model in the present embodiment.

  As shown in the example of teacher data in FIG. 5A, the time prediction model uses the one-hour integrated load power amount for 3 hours immediately before the time when the prediction is performed as an input parameter, and uses the hour of the prediction time as an output parameter. . When the neuro model selected by the neuro model selection unit 17 is a time prediction model, the neuro control unit 182 creates a weighting file as a learning result of the neuro for all the times in the control time zone, and each weighting file. Is updated by learning the load power amount for the past four weeks accumulated in the power load history storage unit 10. Note that the prediction is performed when the time is approaching the control time zone in which the neuro control is performed, and the power amount is sequentially predicted every hour. This time prediction model is effective when living with a pattern at a certain time, although there are fluctuations as a whole.

  As shown in the example of the teacher data in FIG. 5B, the fluctuation prediction model uses the power load rising time, the one-hour integrated load energy at the power load rising time, and the rising gradient as input parameters. The 1 hour integrated load power amount after 1 hour from the rise time of the power load is used as an output parameter. When the neuro model selected by the neuro model selection unit 17 is a fluctuation prediction model, the neuro control unit 182 learns a weighted file that is a neuro learning result from the power history at all times in the control time zone. In the example of the teacher data in FIG. 5B, for example, the rising time of a certain day is 5:00, the load power amount at that time is 0.3 kWh, the rising gradient is 0.1 kWh, and the load after that The amount of power is 0.8 kWh. Learning is performed by calling the teacher data from the control time zone of the power history of the past 4 weeks at the end of the daily control time zone, and the next day, when the time is in the control time zone, the rise of power is recognized. Perform predictions if This fluctuation prediction model is effective when living in a similar pattern, but slightly before and after that time.

  Thus, when the method for controlling the amount of power generation is a control method using a neural network, a neural network neuro model is selected based on the coefficient of variation. One neuro model can be selected from the models.

  Moreover, since the combination of the input parameter and the output parameter for constructing the neuro model is selected, the input parameter and the output parameter can be changed according to the neuro model.

  Next, the criterion of the variation coefficient when the control method determination unit 13 selects arithmetic average control, follow-up control, and neuro control will be described.

  FIG. 6 is a diagram illustrating the relationship between the control method and the coefficient of variation in the present embodiment. As shown in FIG. 6, when the variation coefficient value of the rise time calculated by the variation coefficient calculation unit 12 is 0 to 8 and the variation coefficient value of the peak time is 0 to 5, the control method is determined. The unit 13 determines the method of controlling the power generation amount as arithmetic average control. Further, the fluctuation coefficient value at the power rise time calculated by the variation coefficient calculation unit 12 is 16 or more, the fluctuation coefficient value at the power peak time is 10 or more, and the fluctuation of the peak power amount. When the value of the coefficient is 30 or more and the value of the variation coefficient of the rising slope is 60 or more, the control method determination unit 13 determines the method for controlling the power generation amount to follow-up control.

  As a selection criterion for the time prediction model and the fluctuation prediction model in the neuro control, first, the one having the larger number that falls within the numerical range shown in FIG. 6 is selected. Select The priority item is an item indicated by an underline in FIG. 6. The priority item of the time prediction model is a variation coefficient of each of the rising time and the peak time. The priority item of the variation prediction model is the rising slope and the peak power amount. And the respective coefficient of variation. Further, when the number of priority items is the same, the time prediction model is selected as a default.

  When the predicted power amount is obtained, the neuro control unit 182 transmits a value to the fuel cell 1 as the required power generation amount at the next time. When the predicted load power amount exceeds the reference power amount at the time of power generation, the fuel cell 1 prepares for starting, and when the predicted load power amount is lower than the reference power amount at the time of stop, the fuel cell 1 stops. Can be timed.

  As described above, the history of the power load record is accumulated, the method for controlling the power generation amount is determined from the power load pattern based on the accumulated power load history, and the power generation control is performed according to the determined control method. Depending on the household power load pattern, power generation control is performed using a conventional control method, power generation control is performed using an arithmetic average of past power load results, or power generation is predicted from past power load results. It is determined whether to perform power generation control. Therefore, it is possible to cope with different power load patterns in individual households, and energy saving can be improved. In addition, by learning the load pattern from the power load history and selecting the control method and load prediction method according to each load pattern, it is possible to start and stop the fuel cell according to the load pattern, saving energy and economy. Can be increased. In addition, by learning the daily load pattern, it is possible to sufficiently cope with the seasonal load pattern variation in the home and the family lifestyle itself.

  Since the power load history is acquired, a variation coefficient representing the degree of variation of the power load pattern is calculated, and a control method is determined based on the calculated variation coefficient, so the degree of variation of the power load pattern is represented by a numerical value called a variation coefficient Thus, the control method can be easily determined, and the most suitable control method can be accurately determined.

  In addition, the power load history is acquired, the rise time when the power load increases, the rise gradient, the time when the peak of the power load appears, and the variation coefficient indicating the degree of variation of the power amount at the peak of the power load are calculated and calculated. Since the control method for controlling the power generation amount is determined from one or a plurality of combinations of the variation coefficients, it is possible to reliably select the control method according to the load pattern for each household.

  Furthermore, by the control method determination unit 13, a control method that uses the arithmetic average of the power load amount for each time, a control method that uses a neural network, a control method that makes the start state or the stop state over the entire control time zone, From the power load history, the control method is determined as one control method from among the control methods for selecting one of the start continuation and the stop continuation according to the ratio of the time exceeding the reference power amount to one day. Therefore, a plurality of control methods can be used properly according to the power usage status of each household.

  In this embodiment, a variation coefficient indicating the degree of variation is used as a reference for selecting a control method. However, the present invention is not particularly limited to this, and the correlation coefficient between each factor and the prediction accuracy is selected. It is good also as a standard. In the present embodiment, the neuro model selected by the neuro model selection unit 17 from the neuro model storage unit 19 uses a power integration amount obtained by integrating the power amount for one hour, but the present invention is particularly limited to this. Instead, the temperature and day of the week factors may be used. Furthermore, in order to avoid shortening the learning time or falling into an unexpected prediction result, the weighting file itself learned every certain amount of time may be prepared in advance and selected by the neuro model selection unit 17. In the present embodiment, the control time zone divided by the control time zone dividing unit 11 is divided according to the peak of the power load. However, the present invention is not particularly limited to this, for example, morning, daytime You may simply divide it into 3 nights. Furthermore, it may be divided by a time zone that exceeds a certain load and a time zone that does not exceed a certain load, such as the maximum power generation amount of the fuel cell.

(Second Embodiment)
Hereinafter, a second embodiment of the present invention will be described with reference to the drawings. In the first embodiment described above, the control method determination unit 13 determines one control method from among constant control, arithmetic average control, follow-up control, and neuro control. However, in the second embodiment, the arithmetic average Either one of the control and the neuro control is selected, and the power load is predicted based on the selected prediction method.

  The overall configuration of the fuel cell system according to the second embodiment of the present invention is the same as the overall configuration of the fuel cell system according to the first embodiment shown in FIG.

  FIG. 7 is a block diagram illustrating a configuration of the control unit 6 according to the second embodiment. The control unit 6 includes a power load accumulation unit 8, a calendar information output unit 26, an exception date determination unit 9, a load prediction unit 24, and a power generation control unit 25. In the control unit 6, an initial operation period for determining a prediction method (hereinafter also referred to as a prediction model) for predicting a power load and a normal operation period in which normal operation is performed after the prediction model is determined are provided. The initial operation period was 28 days. During the initial operation period, load follow-up control is performed to follow the power load.

  First, during the initial operation period, the power load accumulation unit 8 acquires a power load from the power meter 5 every 3 seconds, acquires an outside air temperature from an outside air temperature sensor 20 that measures the temperature of the outside air, and outputs a calendar information output unit. 26, date, day of the week, and time calendar information is acquired. Based on the acquired calendar information, the electric load is integrated into the electric load amount for one hour and the average outside temperature for one hour. The power load amount to which the outside air temperature is added is added to the power load history that is data of the past power load amount. When the initial operation period ends, the power load history is output to the exception date determination unit 9 in order to determine the exception date. Thereafter, the exception date flag for each date on which the power load amount is accumulated is acquired from the exception date determination unit 9, and the power load amount and the outside temperature on the day that is the exception date are deleted from the power load history.

  FIG. 8 shows an example of the power load history. As shown in FIG. 8, for example, the power load amount at 0:00 on Thursday, June 22, 2006 is 418.1 kwh, and the outside air temperature is 25.8 degrees. As described above, the power load accumulation unit 8 averages the value obtained by integrating the power load amount acquired every 3 seconds for one hour (418.1 kwh) and the outside air temperature acquired every three seconds (25. 8 degrees) is stored in association with the calendar information (Thursday, June 22, 2006).

  During the normal operation period, necessary information is added to the power load history according to the hourly power load amount and the prediction model acquired from the prediction model selection unit 23, and the power is supplied to the variation coefficient calculation unit 12 once every 24 hours. Outputs load history. When the prediction model acquired from the prediction model selection unit 23 is an arithmetic average model that averages and predicts past power load amounts, only the power load amount is predicted, and the acquired prediction model predicts using a neural network. In the case of the neuro model to be performed, the past power load amount and the past one-hour average outside air temperature are additionally stored in the power load history. In addition, the power load history is output to the exception date determination unit 9 in order to determine the exception date once a day. When the exception date flag acquired from the exception date determination unit 9 is an exception date, the power load amount and the outside temperature for the last 24 hours are deleted from the power load history. If the exception date flag is not an exception date, no processing is performed on the power load history.

  In addition, the power load accumulation unit 8 outputs a power load history to the variation coefficient calculation unit 12 in order to determine a prediction model. Further, the power load accumulation unit 8 outputs the power load history to the load prediction unit 24 in order to predict the power load amount.

  The exception date determination unit 9 acquires the power load history from the power load storage unit 8 once at the end of the initial operation period and once a day during the normal operation period. In either case, an average power load amount is created by averaging the power load amount at the same time in the past 28 days of the acquired power load history.

  After the end of the initial operation period, the day when the average error at the same time between the average power load amount and the daily power load amount exceeds 60% is determined as an exceptional day. As the determination result, an exception date flag for each date (0: exception date, 1: non-exception date) is output to the power load storage unit 8. During the normal operation period, calculate the error for each hour of the power load amount for the last 24 hours immediately before the average power load amount and the power load history are acquired. If the average error for each time exceeds 60%, the exception date Is determined. The determination result is output to the power load storage unit 8 as an exception date flag (0: exception date, 1: non-exception date).

  The exception date is determined by the error of the average power load amount and the power load history for 24 hours immediately before the acquisition of the power load history. This is not the only one.

The variation coefficient calculation unit 12 acquires a power load history from the power load accumulation unit 8 and calculates a variation coefficient indicating variation in the power load amount. FIG. 9 shows an example of the accumulated power load. FIG. 9A is a graph showing the daily change in the power load amount in a home with little variation, and FIG. 9B is a graph showing the daily change in the power load amount in a home with much variation. is there. 9 (a) and 9 (b) show the transition of the power load amount from 0:00 to 23:00 for the past 5 days, the vertical axis shows the power load amount (kwh), and the horizontal axis shows It represents the time. The variation coefficient calculation unit 12 calculates a variation coefficient indicating the degree of variation at the same time of the power load amount of the acquired power load history. The coefficient of variation is obtained by the above equation (1) as in the first embodiment, the average μ is expressed by the above equation (2), and the variance s ² is expressed by the above equation (3).

  The variation coefficient calculation unit 12 outputs a value obtained by averaging the variation coefficients of the power load amount at the same time to the prediction model selection unit 23 as a variation coefficient. A smaller variation coefficient indicates that there is less variation in power load.

  As described above, when the variation coefficient calculation unit 12 calculates the variation coefficient, the variation coefficient of the power load amount at the same time on a plurality of days is averaged, so that the prediction accuracy can be improved.

  The prediction model selection unit 23 holds two arithmetic average models and a neuro model in advance as prediction models, and selects one of the prediction models according to the variation coefficient calculated by the variation coefficient calculation unit 12. .

  The arithmetic average model is a prediction model in which the power load amount at the same time as the time when the power load amount is predicted is an average of the past 28 days, and the prediction load amount is very small. The calculation cost is low because it is expensive and the calculation logic is easy.

  One neuro model is a prediction model that performs prediction using a neural network. In this embodiment, the power load amount for 24 hours immediately before the time for predicting the input value of the neural network and the time immediately before the time for prediction are estimated. The outside air temperature is used, and the output value is the amount of power load at the time of prediction. Since the calculation using the neural network repeatedly performs calculation, the calculation load is heavy on a microcomputer with little memory. In this embodiment, the input value of the neural network is only the past power load amount and the outside temperature. However, depending on other event-based reasoning, it includes the weather forecast, the power consumption amount and operation information of home appliances, etc. Therefore, a large amount of data must be secured. For these reasons, the calculation cost is higher than the arithmetic average. However, even if there is some variation in load generation, load prediction can be performed with higher accuracy than the arithmetic average model.

  For the selection of the prediction model, the relationship between the variation coefficient and the prediction error of the arithmetic average model and the neuro model is used. FIG. 10 is a diagram illustrating the relationship between the variation coefficient and the error in the arithmetic average model and the neuro model, where the vertical axis represents the error (%) and the horizontal axis represents the variation coefficient. The graph shown in FIG. 10 uses average household data of two adults and two children. As the variation coefficient increases, both the arithmetic average model and the neuro model have larger prediction errors. The error is smaller than the arithmetic mean model regardless of the variation coefficient. In order for prediction control to increase power generation efficiency over conventionally used load following control, the prediction error needs to be 25% or less in view of the rise time and stop time of the fuel cell.

  When obtaining the variation coefficient from the variation coefficient calculating unit 12, the prediction model selecting unit 23 obtains an arithmetic average model when the obtained variation coefficient is less than 40 in order to satisfy an error of 25% or less from the graph shown in FIG. When the variation coefficient is 40 or more, the neuro model is selected as the prediction model. In the case of the home in FIG. 9A, since there is little variation, the arithmetic average model is selected, and in the case of the home in FIG. 9B, since there is much variation, the neuro model is selected. Further, the prediction model selection unit 23 outputs the selected prediction model to the power load accumulation unit 8. The power load accumulation unit 8 deletes the outside air temperature information from the power load history when the prediction model selected by the prediction model selection unit 23 is an arithmetic average model.

  Thus, the outside air temperature is acquired from the outside air temperature sensor 20 that measures the outside air temperature, and the acquired outside air temperature is accumulated together with the amount of power load. Then, when a prediction model using a neural network is selected by the prediction model selection unit 23, a prediction load using a neuro model having the power load amount and the outside air temperature as input values and the power generation amount predicted load amount as an output value is used. A quantity is calculated. Therefore, when the selected model is an arithmetic average model, the power load storage unit 8 does not store information other than the power load amount, so that the calculation cost can be reduced in terms of securing the data capacity.

  As described above, when the calculated variation coefficient value is smaller than the predetermined value, a prediction model that uses the arithmetic average of the power load amount at each time is selected, and the calculated variation coefficient value is larger than the predetermined value. Since the prediction model using the neural network is selected, the prediction model can be switched according to the power load pattern to improve the prediction accuracy.

  According to this method, if the variation coefficient is a certain value or more, when control is performed based on the prediction result, the risk that the actual power load amount and the predicted power load amount are completely different increases. It is also possible to perform simple power load tracking without performing the above.

  The load prediction unit 24 acquires a power load history from the power load storage unit 8 every hour. When the prediction model is an arithmetic average model, the load prediction unit 24 outputs a value obtained by averaging the power load amounts from the power load history up to the past 28 days in the same time zone to the power generation control unit 25 as the predicted load amount. When the prediction model is a neuro model, the load prediction unit 24 performs prediction using the power load amount and the outside temperature for the past 28 days for 24 hours immediately before the time of prediction from the power load history as input data. The predicted load amount is learned from learning data in which the power load amount for the past 28 days of time is output data, and the predicted load amount is predicted using the power load amount for 24 hours immediately before the time when the power load history is acquired as input data. I do. The load prediction unit 24 outputs the predicted load amount to the power generation control unit 25 when the calculation of the predicted load amount is completed.

  Next, the flow of the prediction model determination process performed after the end of the initial operation period will be described with reference to FIG. FIG. 11 is a flowchart for explaining a prediction model determination process performed after the end of the initial operation period. First, in step S110, the power load accumulating unit 8 accumulates the power load acquired from the wattmeter 5 and the outside air temperature sensor 20 and the hourly power load amount and the one hour average outside air temperature over 28 days from the outside air temperature. To go. Next, in step S120, the exception date determination unit 9 determines whether it is an exception date based on the power load history accumulated by the power load accumulation unit 8. In step S130, the variation coefficient calculation unit 12 obtains a variation coefficient indicating the degree of variation at the same time from the accumulated power load, and in step S140, the prediction model selection unit 23 determines a model for performing load prediction.

  Next, the flow of processing during the normal operation period will be described with reference to FIG. FIG. 12 is a flowchart for explaining processing during a normal operation period.

  First, in step S150, the power load accumulation unit 8 determines whether or not the current time is 0:00. When the processing time is 0 o'clock (YES in step S150), in step S151, the exception date determination unit 9 determines whether the power load amount for the last 24 hours is an exception date. When it is determined that it is an exceptional day (YES in step S151), in step S152, the power load storage unit 8 deletes the power load corresponding to the exceptional date from the power load history.

  When the prediction model selected after the initial operation period is a neuro model, in step S160, the power load storage unit 8 integrates the power load acquired from the wattmeter 5 every hour to obtain the power load amount, and the outside air temperature sensor 20 The outside air temperature is acquired from the above and the average outside air temperature for one hour is calculated. Next, in step S170, the power load storage unit 8 additionally stores the calculated power load amount and the outside air temperature in the power load history.

  When the prediction model selected after the initial operation period is an arithmetic average model, in step S160, the power load accumulation unit 8 integrates the power load acquired from the wattmeter 5 every hour to obtain a power load amount, and in step S170. The power load storage unit 8 additionally stores the calculated power load amount in the power load history. At this time, the outside temperature is not calculated and additionally stored.

  Next, in step S180, if the prediction model is a neuro model, the load prediction unit 24 performs learning using the power load history for the past 28 days, and then uses the power load history for 24 hours immediately before the prediction time. The amount of power load at the time of is predicted. Further, if the prediction model is an arithmetic average model, the load prediction unit 24 obtains the average of the power load amount for the past 28 days and sets it as the predicted load amount. Next, in step S <b> 190, the power generation control unit 25 uses the predicted load amount calculated by the load prediction unit 24 to output the power generation amount with the optimum efficiency to the fuel cell 1.

  These processes from step S150 to step S190 are repeated every hour.

  From the above, in the present embodiment, the variation coefficient calculation unit 12 obtains the variation coefficient from the amount of power accumulated in the power load accumulation unit 8, and the prediction model selection unit 23 selects a prediction model corresponding to the variation coefficient. It will be. Therefore, it is possible to select a prediction model according to the variation of the power load pattern for each household, and as a result, it is possible to reduce the calculation cost.

  In this embodiment, the prediction model is selected after the end of the initial operation period, but it is also possible to change the household power load pattern by periodically selecting the model, such as at the turn of the season. Is possible.

  In this embodiment, the prediction target is only the power load, but it is also possible to target the hot water supply load.

(Third embodiment)
Next, a third embodiment of the present invention will be described with reference to the drawings.

  FIG. 13 is a block diagram illustrating a configuration of the control unit 6 in the third embodiment. The overall configuration of the fuel cell system according to the third embodiment of the present invention is the same as the overall configuration of the fuel cell system according to the first embodiment shown in FIG. The control unit 6 includes a power load storage unit 8, a calendar information output unit 26, an exception date determination unit 9, a predicted time zone division unit 27, a load prediction unit 24, and a power generation control unit 25.

  The power load accumulating unit 8 acquires a power load from the wattmeter 5 every 3 seconds, and simultaneously acquires the outside air temperature from the outside air temperature sensor 20 and the calendar information of date, day of the week, and time from the calendar information output unit 26. The power load storage unit 8 integrates the power load with the amount of power load for one hour based on the acquired calendar information, and sets the outside air temperature to an average value for one hour. Then, the power load accumulation unit 8 adds the power load amount for one hour to which the calendar information and the average outside air temperature are added to the power load history that is data of the power load amount for each past hour.

  The power load accumulation unit 8 outputs the power load history to the exception date determination unit 9 only when the time when the power load amount is added to the power load history is the end time in the divided time zone information. The exception date determination unit 9 determines the exception date based on the power load history output from the power load storage unit 8. Thereafter, when the content of the exception date flag acquired from the exception date determination unit 9 is an exception date, the power load accumulation unit 8 deletes all the power load amounts from the date when the power load history is acquired from the power load history, Thereafter, until the date changes, even if the divided time zone is changed, the power load amount on the day determined to be an exceptional day is not added to the power load history. The power load accumulation unit 8 does not perform any process on the power load history when the contents of the exception date flag is not an exception date. Further, the power load accumulation unit 8 outputs the power load history to the prediction time zone division unit 27 in order to divide the day into several time zones according to the load generation state once every 24 hours, for example, at 0:00. . In addition, the power load accumulation unit 8 outputs a power load history to the variation coefficient calculation unit 12 in order to determine a prediction model for each divided time period, and supplies power to the load prediction unit 24 in order to predict the power load amount. Outputs load history.

  The exception date determination unit 9 acquires the power load history from the power load accumulation unit 8 and acquires the divided time zone information from the predicted time zone dividing unit 27. From the acquired power load history, an average power load amount is created by averaging the power load amount at the same time in the past 28 days. For each time of the average power load amount in the time zone to which the time when the power load history is acquired and the power load amount in the time zone to which the time at which the power load history is acquired belongs on the day when the power load history is acquired When the average error for each time exceeds 60%, the day on which the prediction is made is determined as an exceptional day. The determination result is output to the power load storage unit 8 as an exception date flag (0: exception date, 1: non-exception date).

  In this way, the power load history is acquired, and it is determined whether or not the acquired power load history is an exceptional day that is an exceptional load pattern different from other days. Since it is controlled not to store the power load amount on the exception day in the history, it is possible to delete the data that is not appropriate for predicting the power load, to accurately store the power load history, and to improve the prediction accuracy Can do.

  By this processing, for example, the power load history can be accumulated except for exceptional days such as going out on a trip or the like, and the power load can be predicted with little error. Moreover, in this Embodiment, although all the electric power loads of the day which became an exceptional day are removed from an electric power load log | history, you may delete only the time slot | zone used as the exceptional day.

  The exception date is determined by using an error in the averaged power load history. However, there is a method of determining with an occurrence probability of a power load rise time, a gradient, or the like, and the present invention is not limited to this.

  The prediction time zone dividing unit 27 acquires the power load history from the power load accumulation unit 8 and averages the power load history within the past 28 days from the acquired power load history at the same time every 24 hours. The power load amount averaged in the period is divided for each time period from the time when the valley is drawn to the time when the next valley is drawn. A combination of a start time and an end time for each divided time zone is generated as divided time zone information.

  FIG. 14 is a diagram for explaining the divided time zones, and is an example in which one day from 0:00 to 23:00 is divided into a plurality of time zones. In the average power load amount shown in FIG. 14, the time from the falling to the rising is 4 o'clock, 10 o'clock, and 14 o'clock. Therefore, in the example of FIG. 14, 24 hours a day is divided from 0:00 to 4 o'clock, 4 o'clock to 10 o'clock, 10 o'clock to 14 o'clock, and 14:00 to 0 o'clock. In this embodiment, the time zone is divided using an averaged power load, but a method of dividing according to the size of a variation coefficient described later is also conceivable.

  The prediction time zone division unit 27 outputs the obtained division time zone information to the variation coefficient calculation unit 12 in order to determine the prediction model of the power load for each division time zone. The predicted time zone dividing unit 27 also outputs the divided time zone information to the exception date determining unit 9 for exception day determination.

  The variation coefficient calculation unit 12 acquires the division time zone information from the prediction time zone division unit 27 and the power load history from the power load accumulation unit 8, and calculates the variation coefficient. The prediction model selection unit 23 selects a prediction model for each divided time zone based on the calculated variation coefficient. Then, the prediction model selection unit 23 outputs, to the load prediction unit 24, prediction model information for each divided time zone that associates the start time and end time of the divided time zone with the prediction model for each divided time zone.

FIG. 15 shows a supplementary diagram for explaining the variation coefficient calculation method. The variation coefficient calculation unit 12 acquires the power load history from the power load accumulation unit 8 and acquires the divided time zone information from the predicted time zone dividing unit 27. A variation coefficient indicating the degree of variation is calculated at the same time for 28 days from the amount of power load in the acquired power load history. The coefficient of variation is obtained by the above equation (1) as in the first embodiment, the average μ is expressed by the above equation (2), and the variance s ² is expressed by the above equation (3).

  Next, a value obtained by averaging the variation coefficient of the power load amount at the same time for each divided time period is set as a variation coefficient for the divided time period. The variation coefficient calculation unit 12 obtains one variation coefficient for each divided time period. For example, in FIG. 15, since the time zone of the divided time zone D2 is from 4 o'clock to 10 o'clock, the average from the variation coefficient Cv (4) to Cv (9) at each time of the divided time zone D2 is divided. The variation coefficient of the time zone D2 is used. A smaller variation coefficient indicates that there is less variation in power load.

  The prediction model selection unit 23 holds an arithmetic average model and a neuro model as prediction models in advance, and selects one of the prediction models for each divided time period. The arithmetic average model is a prediction model that uses the average of the power load amount at the same time as the time when the power load amount is predicted for the past 28 days as the predicted load amount, and predicts when the load variation is very small. It is characterized by low calculation cost due to high accuracy and easy calculation logic.

  One neuro model is a prediction model that performs prediction using a neural network. In this embodiment, the power load amount for 24 hours immediately before the time when the input value of the neural network is predicted and the outside temperature at the time when the prediction is performed. And the output value is the amount of power load at the time of prediction. Since calculation using a neural network repeatedly performs calculation, a long calculation time is required on a microcomputer with little memory. In addition, because fluctuations in temperature affect household loads, there are calculation conditions that require the acquisition of outside air temperature data in addition to the power load history, and the calculation cost is higher than the arithmetic average. There is. However, even if there is variation in load generation, load prediction can be performed with higher accuracy than the arithmetic average model.

  For the selection of the prediction model, the relation between the variation coefficient and the error in the arithmetic average model and the neuro model is used. Note that the selection of the prediction model is the same as that in the second embodiment, and a description thereof will be omitted.

  When the prediction model selection unit 23 acquires the variation coefficient from the variation coefficient calculation unit 12 and the division time zone information from the prediction time zone division unit 27, the obtained variation coefficient is less than 40 in order to satisfy an error of 25% or less from the graph. In this case, the arithmetic average model is selected as the prediction model, and when the obtained variation coefficient is 40 or more, the neuro model is selected as the prediction model. The prediction model information for each divided time zone in which the selected prediction model and the divided time zone are associated one-to-one is generated. FIG. 16 shows an example of prediction model information for each divided time zone. As shown in FIG. 16, the prediction model information for each divided time zone is associated with a prediction model for each divided time zone. For example, a division time zone D1 whose start time is 0 o'clock and end time is 4 o'clock is associated with an arithmetic average model, and the start time is 4 o'clock and the end time is 10 o'clock An arithmetic average model is associated with the time zone D2, a neuro time model is associated with the divided time zone D3 having a start time of 10:00 and an end time of 14:00, and the start time is A neuro model is associated with the divided time zone D4 at 14:00 and the end time of 0 o'clock.

  In this way, in the acquired power load history, the power load amount for a certain period is averaged at the same time, and the averaged power load amount at each time changes from the falling to the rising time to the next. The period from the falling to the rising time is divided as one divided time zone.

  Thereby, the prediction model can be selected according to the variation coefficient of the power load so as to be within the error of the control condition.

  According to this method, if the variation coefficient is a certain value or more, when control is performed based on the prediction result, the risk that the actual power load amount and the predicted power load amount are completely different increases. It is also possible to make a determination to perform simple power load tracking control without performing the above.

  When the load prediction unit 24 acquires the power load history from the power load accumulation unit 8, the load prediction unit 24 acquires the prediction model information for each divided time zone from the prediction model selection unit 23. When the prediction model is an arithmetic average model, the load prediction unit 24 outputs a value obtained by averaging the power load amounts from the power load history up to the past 28 days in the same time zone to the power generation control unit 25 as the predicted load amount. When the prediction model is a neuro model, the load prediction unit 24 uses the power load amount of the past 28 days for 24 hours immediately before the time of prediction from the power load history as input data, and the past 28 days of the time of prediction. The predicted load amount is learned from the learning data using the power load amount for the minute as output data, and the predicted load amount is predicted using the power load amount for 24 hours immediately before the time when the power load history is acquired as input data. The load prediction unit 24 outputs the predicted load amount to the power generation control unit 25 when the calculation of the predicted load amount is completed.

  When the power generation control unit 25 acquires the predicted load amount from the load prediction unit 24, the power generation control unit 25 outputs the power generation amount with the optimum efficiency to the fuel cell 1 using the predicted load amount.

  Next, the flow of processing will be described using the flowchart of the embodiment of the present invention shown in FIG. FIG. 17 is a flowchart for explaining the process during the normal operation period in the third embodiment, and shows the process when the time is on the hour. In addition to the processing on the hour of FIG. 17, the power load storage unit 8 continues to acquire the power load, calendar information, and outside temperature every 3 seconds.

  First, in step S210, the power load accumulation unit 8 determines whether or not the current time is 0:00. When the time is 0 o'clock (YES in step S210), the exception date determination unit 9 sets the exception date flag to a non-exception date “1” (step S211). If the time is not 0:00 (NO in step S210), the process proceeds to step S220.

  Next, in step S212, the prediction time zone dividing unit 27 acquires the power load history from the power load storage unit 8, and averages the power load history within the past 28 days from the acquired power load history at the same time. The power load amount averaged for 24 hours every hour is divided into one day for each time period from the time when the valley is drawn to the time when the next valley is drawn. Then, the predicted time zone division unit 27 outputs the division time zone information to the variation coefficient calculation unit 12.

  Next, in step S213, the variation coefficient calculation unit 12 uses the division time zone information acquired from the prediction time zone division unit 27 and the power load history acquired from the power load accumulation unit 8 for each division time zone. Obtain the variation coefficient of.

  Furthermore, in step 214, the prediction model selection unit 23 selects a prediction model for each divided time period according to the variation coefficient for each divided time period calculated by the variation coefficient calculating unit 12, and creates prediction model information for each divided time period. .

  Next, in step S220, the exception date determination unit 9 determines whether the exception date flag is “0”, that is, whether the date when the power load is acquired is an exception date. If (YES in step S220), the process proceeds to step S250 without accumulating the power load history. If it is not an exceptional day (NO in step S220), in step S230, the power load accumulation unit 8 integrates the acquired power load into the power load amount for one hour. Next, in step S240, the power load accumulating unit 8 adds 1 hour average outside air temperature information, adds to the power load history, and holds it.

  Next, in step S250, the exception date determination unit 9 determines whether or not the time matches the end time included in the divided time zone information. Here, when it is confirmed that the time coincides with the end time included in the divided time zone information (YES in step S250), in step S251, the exception date determination unit 9 starts from the power load accumulation unit 8. The exceptional date is determined based on the acquired power load history and the divided time zone information acquired from the predicted time zone dividing unit 27. If it is determined that the time does not coincide with the end time included in the divided time zone information (NO in step S250), the process proceeds to step S260.

  When it is an exceptional date (YES in step S251), the exceptional date determination unit 9 sets the exceptional date flag to “0” (step S252). Then, the power load accumulation unit 8 deletes all the power load amounts on the exceptional day when the power load is acquired together with the outside air temperature information (step S253). If it is not an exceptional day (NO in step S251), the process proceeds to step S260.

  Next, the process proceeds to step S260 and step S270, which must be performed every hour. In step S <b> 260, the load prediction unit 24 uses the power load history acquired from the power load storage unit 8 and the prediction model information for each divided time period acquired from the prediction model selection unit 23 to predict the predicted load at the next time. Ask for. If the prediction model of the divided time zone that includes the prediction time is a neuro model, learning is performed using the power load history of the past 28 days, and the next is performed using the power load history of 24 hours immediately before the prediction time. The amount of power load at the time of is predicted. In the case of an arithmetic average model, an average of the past 28 days is obtained and used as a predicted load amount. Next, in step 270, the power generation control unit 25 outputs the power generation amount with the optimum efficiency to the fuel cell 1 using the predicted load amount.

  These steps S210 to S270 are repeated every hour.

  From the above, in the present embodiment, the time obtained by dividing one day into a plurality of time zones by the prediction time zone dividing unit 27 and dividing by the variation coefficient calculating unit 12 from the amount of power accumulated in the power load accumulation unit 8. A variation coefficient is obtained for each band, and the prediction model selection unit 23 selects a prediction model corresponding to the variation coefficient. Therefore, it is possible to select a prediction model according to variations in the power load pattern for each time period, and as a result, it is possible to reduce the calculation cost.

  In this embodiment, the timing for calculating the predicted load amount is every hour, but it may be calculated in units of minutes. Moreover, although two types of prediction models, the arithmetic average model and the neuro model, are used, the number of options may be increased by adding types, even if changing to another prediction model.

  In this embodiment, the prediction target is only the power load, but it is also possible to target the hot water supply load.

  According to such a configuration, a variation coefficient is obtained based on the amount of power load in the home in the past, and according to the variation coefficient, a prediction method is a simple prediction method such as arithmetic average or a calculation load such as a neural network. It is possible to select such a prediction method, and when complicated calculation is not required, it is possible to perform prediction while suppressing calculation cost.

  The fuel cell system according to the present invention can cope with different electric power loads for each home, can improve energy saving, and is useful as a fuel cell system for supplying electric power using raw fuel gas. Moreover, it can be applied to uses such as a control method according to a hot water supply load pattern, and can also be applied to prediction of hot water supply loads such as gas engine cogeneration and CO2HP.

1 is a diagram showing an overall configuration of a fuel cell system according to the present invention. It is a block diagram which shows the structure of the control part in 1st Embodiment. It is a figure for demonstrating the control time slot | zone divided | segmented into plurality. It is a figure for demonstrating determination of the exceptional day in this embodiment. It is a figure for demonstrating the input / output parameter of the neuro model in this embodiment. It is a figure which shows the relationship between the control method in this embodiment, and a variation coefficient. It is a block diagram which shows the structure of the control part in 2nd Embodiment. It is a figure which shows an example of the electric power load log | history in this embodiment. It is a figure which shows the past power load amount in a home with little dispersion | variation in power consumption. It is a figure which shows the relationship between the variation coefficient and error in an arithmetic mean model and a neuro model. It is a flowchart for demonstrating the process of the prediction model determination performed after completion | finish of an initial driving | operation period. It is a flowchart for demonstrating the process in a normal driving | operation period. It is a block diagram which shows the structure of the control part in 3rd Embodiment. It is a figure for demonstrating a division | segmentation time slot | zone. It is a figure for demonstrating the variation coefficient calculation method. It is a figure which shows an example of the prediction model information according to division | segmentation time slot | zone. It is a flowchart for demonstrating the process in the normal driving | operation period in 3rd Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Fuel cell 2 Power output line 3 Load 4 Commercial power supply 5 Wattmeter 6 Control part 8 Power load storage part 9 Exception day determination part 10 Power load history storage part 11 Control time zone division part 12 Variation coefficient calculation part 13 Control method determination part DESCRIPTION OF SYMBOLS 14 Constant control part 16 Tracking control part 17 Neuro model selection part 19 Neuro model storage part 20 Outside temperature sensor 23 Prediction model selection part 24 Load prediction part 25 Electric power generation control part 151 Load prediction part 152 Arithmetic average control part 181 Load prediction part 182 Neuro Control unit

Claims (15)

A fuel cell system for supplying power using fuel gas,
Power load storage means for storing a history of power load results;
Control method determining means for determining a control method for controlling the power generation amount from the power load pattern based on the power load history stored by the power load storing means;
A fuel cell system comprising: power generation control means for performing power generation control according to the control method determined by the control method determination means. It further comprises variation coefficient calculation means for obtaining a power load history accumulated by the power load accumulation means and calculating a variation coefficient representing a degree of variation in the power load pattern,
2. The fuel cell system according to claim 1, wherein the control method determining means determines a control method based on the variation coefficient calculated by the variation coefficient calculating means. The variation coefficient calculating means acquires the power load history accumulated by the power load accumulating means, the rise time when the power load increases, the rise gradient when the power load increases, the time when the peak of the power load appears, and the power load Calculate the coefficient of variation indicating the degree of variation in the amount of power at each peak,
3. The fuel cell system according to claim 2, wherein the control method determining means determines a control method from one or a combination of a plurality of variation coefficients calculated by the variation coefficient calculation means.   When the variation coefficient of the rise time at which the power load increases is equal to or less than the predetermined value and the variation coefficient of the power amount at the peak of the power load is equal to or less than the predetermined value, the control method determining means 4. The fuel cell system according to claim 3, wherein the control method uses an arithmetic average of the power load amount.   The control method determining means is a time at which the variation coefficient of the rise time at which the power load increases is equal to or greater than a predetermined value, the variation coefficient of the rise gradient at which the power load increases is equal to or greater than a predetermined value, and the peak of the power load appears When the variation coefficient of the power amount is equal to or greater than a predetermined value and the variation coefficient of the electric energy at the peak of the power load is equal to or greater than the predetermined value, the control method is determined to follow a past power load pattern from the power load history. The fuel cell system according to claim 3.   6. The fuel cell according to claim 3, wherein the control method determining means determines a control method using a neural network when the conditions described in claim 4 and claim 5 are not satisfied. system.   When the control method determined by the control method determination unit is a control method using a neural network, a neuro model selection unit that selects a neuro model of the neural network based on the variation coefficient calculated by the variation coefficient calculation unit The fuel cell system according to claim 6, further comprising:   8. The fuel cell system according to claim 7, wherein the neuro model selection means selects a combination of an input parameter and an output parameter for constructing the neuro model according to the neuro model. The variation coefficient calculating means calculates a variation coefficient representing a variation degree of the power load amount at the same time in a certain period,
When the variation coefficient calculated by the variation coefficient calculation unit is less than a predetermined value, the control method determination unit determines a control method using an arithmetic average of the power load amount at each time from the power load history, and calculates the variation coefficient If the variation coefficient calculated by the means is greater than or equal to a predetermined value, determine the control method using a neural network,
3. The fuel according to claim 2, wherein the power generation control means calculates a predicted load amount using the control method determined by the control method determination means, and performs power generation control according to the calculated predicted load amount. Battery system. The power load history accumulating means acquires the outside air temperature from an outside air temperature sensor that measures the outside air temperature, accumulates the acquired outside air temperature together with the amount of power load,
When the power generation control means is determined to be a control method using a neural network by the control method determination means, the power load amount and the outside air temperature are input values, and the power generation amount predicted load amount is a neuron having an output value. 10. The fuel cell system according to claim 9, wherein a predicted load amount is calculated using a model, and power generation control is performed according to the calculated predicted load amount. The power load history accumulated by the power load history accumulating means is acquired, and one day is divided into a plurality of divided time zones, and divided time zone information which is a combination of a start time and an end time for each divided time zone Further comprising a time zone dividing means for generating
The variation coefficient calculating means acquires the power load history accumulated by the power load accumulation means and the divided time zone information generated by the time zone dividing means, and the power load for each divided time zone. Calculate a variation coefficient that represents the degree of variation in quantity,
The control method determining means acquires the variation coefficient calculated by the variation coefficient calculating means and the divided time zone information generated by the predicted time dividing means, and for each divided time zone based on the variation coefficient Determining the control method,
The power generation control unit calculates a predicted load amount for each of the divided time zones using the control method determined by the control method determination unit, and performs power generation control according to the calculated predicted load amount. The fuel cell system according to claim 9.   The time zone dividing unit averages the power load amount for a certain period of the power load history acquired from the power load accumulation unit at the same time, and the averaged power load amount at each time is established. 12. The fuel cell system according to claim 11, wherein a time period from a change from falling to rising to a time from the next falling to rising is divided into one divided time zone.   The control method determining means determines the control method to select either one of start and stop according to the ratio of the time exceeding the reference power amount in the divided time zone from the power load history. Item 12. The fuel cell system according to Item 11.   10. The fuel cell system according to claim 9, wherein the variation coefficient calculation means calculates the variation coefficient of the power load amount at the same time on a plurality of days when calculating the variation coefficient.   The power load history accumulated by the power load history accumulating means is acquired, and it is determined whether the acquired power load history is an exceptional day that is an exceptional load pattern different from other days, and an exception 15. The method according to claim 1, further comprising: an exception date determination unit that performs control so as not to accumulate the power load amount of the exception day in the power load history accumulated by the power load accumulation unit. A fuel cell system according to claim 1.

Images