JP6641849B2 - Power prediction method - Google Patents

Description

  The present invention relates to a power prediction method for predicting a total power demand in a factory.

  Factories that use a lot of power, such as steelworks, purchase power by signing a power purchase contract with a power company. Generally, in a power purchase contract, the contract power is determined based on the maximum power consumption (hereinafter, also referred to as a “demand value”) for a predetermined time (for example, every 30 minutes), and the actual demand value is determined by the contract power. Penalties will be imposed if you exceed. On the other hand, even when the demand value is significantly lower than the contracted power, the amount of the contracted power is paid. Under such a contract, it is economically important on the factory side to adjust the demand so that the power consumption is always kept stable, and therefore, it is necessary to accurately forecast the power demand of the factory. There is a need.

  As a method of estimating power demand, in the past, as described in Patent Literature 1, it is necessary to obtain past actual power consumption for a predetermined time and estimate the power demand for a predetermined time in the future from the value of this power consumption. There is a method based on series analysis. On the other hand, for example, Patent Literatures 2 and 3 relate to estimating the power demand of a rolling mill for steel sheets, and associate the power demand with the specifications or related operation parameters of the steel sheet to be processed according to the rolling processing schedule by mathematical expressions. Accordingly, a prediction method based on an operation plan of a factory for estimating power demand has been proposed.

JP-A-55-136832 JP-A-6-262223 JP 2001-321810 A

Genshiro Kitagawa, Introduction to Time Series Analysis, Iwanami Shoten, 2005 E. Primiceri, “Time Varying Structural Vector Autoregressions and Monetary Policy”, Reviews of Economic Studies 72, 2005, p.821 Takafumi Tanaka, "Introduction to Time Series Analysis with R", CAB Publishing, 2008

  However, as described in Patent Literature 1, the power prediction accuracy based on the time series analysis using the past actual values is not very high, and the predicted values and the actual values may greatly deviate. The reason is that the power consumption is not based on a fixed chronological rule but based on an artificial operation plan, and it is considered that this plan is also likely to be frequently changed.

  Also, the methods described in Patent Documents 2 and 3 also include uncertainty due to a change in the plan. In addition, even if the operations are based on the same operation parameters, the same power is not necessarily consumed, and it is not easy to obtain a highly accurate relational expression that relates the power demand to various operation parameters.

  Therefore, the present invention has been made in view of the above problems, and it is an object of the present invention to provide a new and improved power prediction method capable of improving the prediction accuracy of power demand. It is in.

  According to one aspect of the present invention, there is provided a power prediction method for predicting a total power demand of a factory, the method comprising: calculating a power demand based on a time-series prediction model using actual data on power consumption. A first probability density function calculating step of calculating a first probability density function for estimating the power demand, and a second probability density function of calculating a second probability density function for estimating the power demand based on the operation plan of the factory A calculating step, a third probability density function calculating step of calculating a third probability density function represented based on a product of the first probability density function and the second probability density function, and a third probability density function And a power estimation step of determining a representative value of the power estimation value as a power estimation value.

  In the power prediction step, one of the mode, the average, and the median may be determined as the representative value of the third probability density function.

  In the power prediction step, a prediction error of the third probability density function may be calculated.

  As described above, according to the present invention, it is possible to improve the prediction accuracy of power demand.

It is an explanatory view showing the functional composition of the power prediction device concerning one embodiment of the present invention. It is a flow chart which shows an example of the power prediction method concerning the embodiment. 4 is a graph showing an example of a probability density function of a predicted value of power in Comparative Example 1, Comparative Example 2, and Example 1 of the present invention.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the specification and the drawings, components having substantially the same functional configuration are denoted by the same reference numerals, and redundant description is omitted.

<1. Concept explanation>
The present inventors diligently studied a technique for improving the prediction accuracy of power demand. As a result, prediction by a time-series model using actual data (hereinafter, referred to as “actual value prediction”) and prediction by an empirical formula based on an operation plan (hereinafter, referred to as “planned value prediction”) are: Since they are calculated based on completely different prediction principles using completely different data, they are considered to be stochastically independent of each other. That is, the inventors of the present application considered the two predictions in a probabilistic manner, and thought that by fulfilling the condition that the two predictions are simultaneously satisfied, the accuracy of the prediction can be improved.

  In the present invention, the prediction at the future time t is not considered as one power value but as a probability density function using the power value e as a variable. Assuming that the probability density function of actual value prediction at time t is P1 (e) and the probability density function of plan value prediction is P2 (e), the probability density function P3 (e) that is satisfied by two independent predictions at the same time is For example, it can be expressed as the following equation (1). The probability density function P3 (e) of the equation (1) calculates the product of the probability density function P1 (e) and the probability density function P2 (e), and further calculates the probability density function P1 (e) and the probability density function P2 ( This is a value obtained by dividing the product with e) by a value obtained by integrating e over the entire interval.

  The prediction according to the present invention can be stochastically represented by the probability density function P3 (e) calculated by the equation (1), and a representative value (for example, a mode value; The power value e) that gives the maximum value of is the power prediction value. Further, the reliability of the predicted power value can be calculated from the probability density function P3 (e). Since the denominator of equation (1) is a constant, the value of the denominator does not affect the mode of the probability density function P3 (e). Therefore, it is possible to calculate the predicted power value only from the product of the probability density function P1 (e) and the probability density function P2 (e) without calculating the value of the denominator of the equation (1).

Here, the probability density function P1 (e) of the actual value prediction is obtained from the characteristic parameters of the prediction model determined by applying the past power consumption actual data to the time-series prediction model and the actual power value immediately before the prediction time. Can be calculated. Many time-series prediction models are known, as introduced in Non-Patent Documents 1 to 3, for example. In many of these models, the standard deviation of the error is calculated together with the prediction value. Is done. And E ₁ the prediction value obtained here, when the standard deviation and sigma _1, the probability density function P1 (e) can be represented as a normal distribution function as shown in Equation (2).

On the other hand, the probability density function P2 (e) of the plan value prediction can be set by the plan prediction value of the power calculated based on some process parameters determined based on the production plan value and the prediction error. . Here, the plan prediction value can be obtained, for example, by substituting the value of the process parameter into a regression equation obtained by performing a multiple regression analysis on the relationship between the past process parameter and the power consumption. As the prediction error, the power estimation error in the multiple regression equation may be used. Assuming that a predicted value obtained by multiple regression analysis from each process parameter is E ₂ and the standard deviation of the error is σ ₂ , the probability density function P2 (e) is expressed by the following equation (3).

  By applying Equations (2) and (3) to Equation (1), the probability density function P3 (e) can be specifically calculated. Since the mode value of the calculated probability density function P3 (e) is the power value with the highest realization probability, this value can be used as the predicted value. Here, the value used as the predicted value is not limited to the mode, and the average or median of the probability density function P3 (e) may be used as the representative value depending on the situation. The reliability of the prediction can be evaluated by calculating the quantile of the probability density function P3 (e), for example, in a 5% -95% confidence interval.

  When a prediction model based on the Monte Carlo method or the like or a nonlinear multiple regression analysis or the like is used, the probability density functions P1 (e) and P2 (e) may not be able to be expressed as a specific function expression such as a normal distribution function. . However, even in such a case, if the probability density function is obtained in the form of numerical data, the probability density function P3 (e) can be calculated numerically based on the equation (1), and can be used without any problem. it can.

  Further, in the present invention, when calculating the probability density functions P1 (e) and P2 (e), there are no particular restrictions on the time series prediction model to be used, the method of multiple regression analysis, and the method of selecting operation parameters. These can be appropriately selected according to the environment, cost, and the like regarding data collection and calculation so that the error in the actual value prediction and the error in the plan value prediction become small.

  As described above, according to the power prediction method of the present invention, by simultaneously considering the actual value prediction by the time-series model using the actual data and the plan value prediction by the empirical formula based on the operation plan, Demand forecasting accuracy can be greatly improved. As a result, it is possible to significantly improve the accuracy of forecasting the power demands of factories that fluctuate irregularly, and to detect in advance that power demand will be out of the contracted power conditions, and to take protective measures. Time can be obtained. Hereinafter, the power prediction method of the present invention will be specifically described.

<2. Power demand forecast>
A configuration of a power prediction device 100 according to an embodiment of the present invention and a power prediction method based on the configuration will be described. In the following description, as a specific example, power prediction of a rolling mill that performs hot rolling of a steel sheet will be described. In this rolling factory, a slab-shaped steel ingot is heated in a heating furnace, and is rolled into a long strip shape by a series of rolling mills, thereby producing a hot-rolled coil wound in a coil shape.

[2-1. Functional configuration]
First, a configuration example of a power prediction device 100 according to the present embodiment will be described with reference to FIG. FIG. 1 is an explanatory diagram illustrating a functional configuration of a power prediction device 100 according to the present embodiment. The power prediction device 100 is realized by, for example, a computer including a CPU functioning as an arithmetic processing device and a control device. The CPU of the power prediction device 100 controls the entire operation in the power prediction device 100 or a part thereof in accordance with various programs recorded in a ROM, a RAM, a storage device, a removable recording medium, or the like. As shown in FIG. 1, the power prediction device 100 according to the present embodiment includes an actual value acquisition unit 110, a first probability density function calculation unit 120, a plan value acquisition unit 130, and a second probability density function calculation A unit 140, a third probability density function calculating unit 150, and a prediction processing unit 160.

  The performance value acquisition unit 110 obtains performance data including a performance value of power consumption in a factory. In the factory, the power consumed in the entire factory is managed, and the amount of power consumed in each facility in the factory is recorded in the actual data storage unit 10 at a predetermined timing (for example, every 30 minutes). The performance value acquisition unit 110 obtains the power consumption of the entire factory from the performance data storage unit 10 from the performance data storage unit 10 and outputs the power consumption to the first probability density function calculation unit 120.

The first probability density function calculating unit 120 calculates a first probability of predicting power demand from a time-series prediction model of power based on the actual value of power consumption contained in the actual data acquired by the actual value acquiring unit 110. Calculate the density function. As the time series prediction model, for example, an ARIMA (Autoregressive Integrated Moving Average) model, a BVAR-TP model (Bayesian / time-varying parameter / autoregression model), or the like can be used. First probability density function calculation unit 120, for example, than the actual value of the past power consumption back a predetermined period from the present, in the next period by using a time series forecast model of power power prediction value E ₁ and the error of of calculating a standard deviation sigma _1, calculates a first probability density function based on the equation (2) P1 (e).

  The plan value acquisition unit 130 acquires an operation schedule set as a production plan. In the factory, a production plan is managed by a system, and a daily operation schedule is stored in the plan storage unit 20. Past operation results are also recorded in the plan storage unit 20, and if there is a change in the operation schedule, it is appropriately reflected. Further, in order to estimate the power demand from the operation schedule, for example, the processing length Lph (m / h) and the processing weight Wph (Kg / h) of the coil per hour are calculated based on the production plan as the processing amount of the rolling process. The prediction is made up to a predetermined period ahead, and the predicted processing amount is also stored in the plan storage unit 20. The plan value acquisition unit 130 acquires the predicted processing amount in the next period from the plan storage unit 20 and outputs the same to the second probability density function calculation unit 140.

Second probability density function calculation section 140 calculates the predicted value E ₂ of the power demand on the basis of the predicted processing amount is obtained operation plan by planned value acquisition unit 130, the power demand based on the above formula (3) Is calculated, a second probability density function P2 (e) for estimating. Power demand prediction value E ₂ is calculated for example by the following equation (4). The constants C0, C1, C2, and C3 are determined in advance by, for example, a linear multiple regression analysis based on the past power consumption values, the coil processing length Lph, and the processing weight Wph. Further, the standard deviation σ ₂ of the error of the predicted value E ₂ of the power demand is also calculated in advance by the linear multiple regression analysis.

  The third probability density function calculator 150 calculates a third probability density function P3 (e) based on the product of the first probability density function P1 (e) and the second probability density function. The third probability density function P3 (e) is represented, for example, by the above equation (1).

  The prediction processing unit 160 determines a power prediction value in the next period to be predicted based on the third probability density function P3 (e). The power prediction value is a representative value of the third probability density function P3 (e), and can be, for example, a mode, an average, or a median. Further, the prediction processing unit 160 calculates a prediction error of the predicted power value. As the prediction error of the predicted power value, the prediction processing unit 160 calculates, for example, the 5% quantile and the 95% quantile of the third probability density function P3 (e). The prediction processing unit 160 outputs the calculated power prediction value and the prediction error to the output unit 30 as prediction information. The output unit 30 is, for example, a display device such as a display, and notifies the prediction information input from the prediction processing unit 160 so that the factory manager can refer to the prediction information.

[2-2. Power prediction method]
Next, a power prediction method by the power prediction device 100 according to the present embodiment will be described with reference to FIG. FIG. 2 is a flowchart illustrating an example of the power prediction method according to the present embodiment.

  In the power prediction method according to the present embodiment, as shown in FIG. 2, first, an actual value prediction process (S100 to S104) and a plan value prediction process (S110 to S114) are executed. These processes may be executed in parallel as shown in FIG. 2, or each process may be executed in order.

(Actual value prediction processing)
First, the actual value prediction process will be described. First, as shown in FIG. 2, the actual value of the power consumption is acquired by the actual value acquisition unit 110 from the actual data storage unit 10 (S100). In the performance data storage unit 10, the amount of power consumed by each facility in the factory is recorded at a predetermined timing. In this embodiment, the amount of power consumed by each facility in the factory is recorded in the actual data storage unit 10 at a timing of once every 30 minutes, where one step is 30 minutes. The actual value acquisition unit 110 acquires the actual value of the power consumption for a predetermined past period set in advance and outputs the acquired actual value to the first probability density function calculating unit 120. For example, the actual value acquisition unit 110 acquires the actual power consumption data 480 steps before (i.e., 10 days before) from the current time, and outputs the acquired power consumption data to the first probability density function calculation unit 120. The number of actual data obtained by actual value acquisition unit 110 is not limited to such an example, it is better large in order to obtain the standard deviation sigma ₁ highly accurate prediction error in step S102 described later.

Next, the first probability density function calculator 120 calculates the predicted value of the power consumption in the period to be predicted based on the actual value of the power consumption acquired in step S100 and the time-series prediction model of the power. E ₁ and calculating the standard deviation sigma ₁ of the prediction error (S102). Here, an ARIMA model is used as a power time-series prediction model (see Non-Patent Document 3). The parameter value of the ARIMA model is determined by, for example, analyzing the actual power consumption data for the past three months using a third-order ARIMA model in advance. The first probability density function calculator 120 calculates the predicted value E ₁ and the standard deviation sigma ₁ for each step of the period of the prediction target from the analysis by the ARIMA model. The first probability density function calculation section 120, based on the predicted values _{E 1} and the standard deviation sigma _1, calculates a than the first probability density function P1 (e) the formula (2) (S104). The first probability density function P1 (e) is calculated for the power value e at a predetermined step size, for example, at 10 KW, and is obtained as an array value.

(Plan value forecasting process)
Next, the plan value prediction processing will be described. In the plan value prediction processing, first, as shown in FIG. 2, the plan value acquisition unit 130 acquires plan values of predetermined process parameters from the operation schedule stored in the plan storage unit 20 (S110). The plan storage unit 20 stores, for example, the processing length Lph (m / h) and the processing weight Wph (Kg / h) of the coil per hour as the past operation results and the processing amount of the rolling process for estimating the power demand. ) Is stored. In the present embodiment, the predicted value of the power is calculated by using the plan values of these two process parameters. However, the process parameters that can be used in the plan value prediction process can be determined as appropriate, and the power consumption in the operation is large. It is advisable to select the process parameters that influence. In the rolling step, in addition to the above, for example, a steel type to be rolled may be used as a process parameter. Further, in the present embodiment, the predicted power value is calculated using two process parameters. However, the present invention is not limited to such an example, and the predicted power value may be calculated using one or a plurality of process parameters. Is possible.

  As the predicted value of the processing amount of the rolling process, the processing amount up to, for example, 6 hours ahead, that is, up to 12 steps ahead is recorded based on the production plan. The plan value acquisition unit 130 acquires the processing amount of the rolling process at least one step ahead in the period to be predicted, and outputs the acquired processing amount to the second probability density function calculation unit 140.

Then, the second probability density function calculation unit 140, based on the throughput of the rolling process in the period of the prediction target acquired in step S110, the predicted values E ₂ and the prediction error of the power consumption in the period calculating a standard deviation σ ₂ (S112). Second probability density function calculation section 140 calculates the predicted value E ₂ of the electric power based on the equation (4) for each step of the period of the prediction target. For example, if the acquired throughput of the rolling process in step S110 until 12 steps ahead, one step away, 2 steps ahead, ..., 12 steps ahead of the power predicted value E ₂ are calculated . The constants C0, C1, C2, and C3 in the above equation (4) are determined in advance by linear multiple regression analysis from the relationship between the past power consumption value and the actual values of the coil processing length Lph and the processing weight Wph. Is done. At this time, the second probability density function calculator 140 also calculates the standard deviation σ ₂ of the error of the predicted value E ₂ of each power from the linear multiple regression analysis.

Then, the second probability density function calculation unit 140, the predicted value of the calculated power E ₂ from the standard deviation sigma ₂ of the error, the second probability density function P2 based on the predicted value E ₂ (e) a It is calculated by the above equation (3) (S114). The second probability density function P2 (e) is calculated for the power value e in a predetermined step size, for example, in 10 KW steps, and is obtained as an array value.

(Power prediction processing)
When the actual value prediction processing and the plan value prediction processing are executed, the third probability density function calculation unit 150 executes the power prediction processing. In the power prediction process, first, the third probability density function P1 (e) calculated in step S104 and the second probability density function P2 (e) calculated in step S114 are used to calculate a third probability density function. A density function P3 (e) is calculated (S120). The third probability density function calculating unit 150 calculates the third probability density function based on the array value of the first probability density function P1 (e) and the array value of the second probability density function P2 (e) using the above equation (1). Calculate numerically. Thereby, the value of the third probability density function P3 (e) in each step to be predicted is obtained numerically.

  Based on the third probability density function P3 (e) obtained in step S120, the third probability density function calculation unit 150 calculates a representative value of the probability density as a predicted value E of power. Further, the third probability density function calculating unit 150 calculates the 5% quantile and the 95% quantile of the third probability density function P3 (e) as the prediction error of the predicted value E (S122). ). The prediction value E and the prediction error are used as prediction information. After calculating the prediction information in each step of the prediction target, the third probability density function calculation unit 150 outputs the prediction information to the output unit 30 (S124). Thereby, the production manager of the factory can adjust the rolling schedule with reference to the prediction information displayed on the output unit 30, for example. Further, the power manager in the factory region can operate the generator at the related power plant and adjust the power interchange between other regions based on the prediction information.

  Hereinafter, regarding the power demand prediction processing of the rolling mill described in the above embodiment, the effectiveness of the power prediction method according to the present embodiment was verified. Hereinafter, the present invention will be described with reference to examples and comparative examples, but the present invention is not limited to the following examples.

(Comparative Example 1)
Comparative Example 1 uses the BVAR-TP model shown in Non-Patent Document 2 based on the power consumption data of the past three months for the power of the rolling mill described in the above embodiment. It is a prediction of power. The mode of this distribution was used as the power prediction value, and the 5% and 95% quantiles of this distribution were used as the prediction error. The probability density function of the prediction of the electric power after 30 minutes obtained by this prediction model has a bell-shaped distribution as shown in the upper part of FIG. The mode e of the probability density function is 62.19 (MW), the 5% quantile obtained from this distribution is 48.27 (MW), and the 95% quantile is 76.11 (MW). MW). In Table 1 below, the predicted values obtained in 30-minute increments from 30 minutes to 150 minutes later, the 5% quantile that is the lower limit of the 5% -95% confidence interval, and the 95% quantile that is the upper limit , And the width of the confidence interval.

(Comparative Example 2)
Comparative Example 2 predicts the power at the same time as Comparative Example 1 in the same rolling mill as Comparative Example 1. As in Comparative Example 1, the mode of this distribution was used as the power prediction value, and the 5% and 95% quantiles of this distribution were used as the prediction error. Here, based on the power consumption for the past three months, the input weight per hour Wph (Kg / hour) and the input length Lph (m / hour) of the coil, the weight expressed by the above equation (4) is obtained. As a result of obtaining the parameters of the regression model, the multiple regression equation for predicting the predicted value E ₂ (MW) of the electric power from the planned values of the input weight Wph and the input length Lph is as shown in the following equation (5). Standard deviation sigma ₂ error of the predicted value _{E 2} became 9.79 (MW).

In equation (5), by substituting the value of the input weight Wph and turned length Lph are planned at time t, and calculates a predicted value E _2. The probability density function of the power prediction after 30 minutes obtained by this prediction model has a normal distribution with an average value of 58.85 (MW) and a standard deviation of 9.79 (MW) as shown in the center of FIG. . The mode of the normal distribution was equal to the mean, 58.85 (MW), the 5% quantile obtained from this distribution was 39.27 MW, and the 95% quantile was 78.43 MW. . In Table 2 below, the predicted values obtained in 30-minute intervals from 30 minutes to 150 minutes later, the 5% quantile that is the lower limit of the 5% -95% confidence interval, and the 95% quantile that is the upper limit , And the width of the confidence interval. Since the standard deviation σ ₂ of the error by the multiple regression analysis in Comparative Example 2 is a constant independent of time, the width of the confidence interval is 36.16 (MW) at any time.

(Example 1)
In the first embodiment, the power demand prediction processing of the rolling mill described in the above embodiment predicts the power at exactly the same point as in the first and second comparative examples. As in Comparative Examples 1 and 2, the mode of this distribution was used as the power prediction value, and the 5% quantile and the 95% quantile of this distribution were used as the prediction error. The probability density function P3 (e) of the predicted value E of the power in the first embodiment is calculated by using the probability density function P1 (e) at the time t calculated in the first comparative example and the time density t calculated in the second comparative example. And the probability density function of the probability density function P2 (e) is newly calculated based on the above equation (1). As shown in the lower part of FIG. 3, the probability density function P3 (e) of the power prediction after 30 minutes has a bell-shaped distribution with a narrow section and a relatively sharp bell. The mode of the probability density function P3 (e) is 60.61 (MW), the 5% quantile obtained from this distribution is 50.14 (MW), and the 95% quantile is 70 .99 (MW). In Table 3 below, the predicted values obtained in 30-minute intervals from 30 minutes to 150 minutes later, the 5% quantile that is the lower limit of the 5% -95% confidence interval, and the 95% quantile that is the upper limit , And the width of the confidence interval. The width of the confidence interval was significantly narrower than the width of the confidence interval of the predicted values of Comparative Examples 1 and 2 at any time, confirming that the reliability of prediction was greatly improved.

  As described above, the preferred embodiments of the present invention have been described in detail with reference to the accompanying drawings, but the present invention is not limited to such examples. It is obvious that those skilled in the art to which the present invention pertains can conceive various changes or modifications within the scope of the technical idea described in the claims. It is understood that these also belong to the technical scope of the present invention.

REFERENCE SIGNS LIST 10 actual data storage unit 20 plan storage unit 30 output unit 100 power prediction device 110 actual value acquisition unit 120 first probability density function calculation unit 130 plan value acquisition unit 140 second probability density function calculation unit 150 third probability density Function calculation unit 160 Prediction processing unit

Claims (3)

A power forecasting method for forecasting the total power demand of a factory,
A first probability density function calculating step of calculating a first probability density function for predicting power demand based on a time-series prediction model using actual data on power consumption;
A second probability density function calculating step of calculating a second probability density function for predicting power demand based on the operation plan of the factory;
A third probability density function calculating step of calculating a third probability density function represented based on a product of the first probability density function and the second probability density function;
A power prediction step of determining a representative value of the third probability density function as a power prediction value;
And a power prediction method.   The power prediction method according to claim 1, wherein in the power prediction step, one of a mode, an average, and a median is determined as a representative value of the third probability density function. The power prediction method according to claim 1, wherein in the power prediction step, a prediction error of the third probability density function is calculated.

Images