EP3491591A1

EP3491591A1 - Method for predicting consumption demand using an advanced prediction model

Info

Publication number: EP3491591A1
Application number: EP17740416.7A
Authority: EP
Inventors: Philippe CHARPENTIER
Original assignee: Electricite de France SA
Current assignee: Electricite de France SA
Priority date: 2016-07-26
Filing date: 2017-07-24
Publication date: 2019-06-05
Also published as: FR3054703A1; FR3054703B1; WO2018019769A1

Abstract

The invention relates to the management of an electrical power distribution network, by estimating power demand from consumers, comprising the steps: a1) predicting a plurality of link functions, each specific to one type of model and to at least one input parameter (S2); b1) receiving, as input of the computer device, a plurality of respective input parameter values (S1); b2) determining, according to a selected criterion, a classification by order of relevance of the parameters for which the values are received in operation b1) (S6), and; a2) constructing a model based on at least one link function using at least one of the parameters that is the most relevant according to the classification of operation b2), for subsequently implementing step c) (S8; S9).

Description

A method for predicting consumption demand, using an improved prediction model

The invention relates to the management of an electrical energy distribution network, by estimating the energy demand by consumers, in particular with the aim of providing for the erasure of targeted consumers (which amounts to imposing maximum consumption on them), limited to a threshold) in case of peak demand.

More particularly, it seeks to predict the erasable power of a day D + 1 to D + X on a portfolio of consumers (called "customers" below) having subscribed to a power greater than a given threshold (for example 250 kW ) (typically at a tariff with a tariff incentive to voluntary deletion). This forecast is then transmitted to a computing entity optimizing the energy supplier in order to optimize the electricity production from its various sources of production (eg nuclear, hydraulic, photovoltaic and / or fossil energy with generator).

There are few state-of-the-art publications for erase prediction as such, although several power consumption prediction techniques are presented in the state of the art. These consumption forecasting techniques highlight solutions where a single model is generally used and model learning is performed on fairly long data histories. In addition, once learning is done, the parameters are programmed hard (permanently) in a computer system consumption forecast / decision making erasure.

It is therefore always the same variables (or variable delays) that are used to make the prediction, and the function called "link function" between the monitored parameters and the consumption forecast is always fixed. Usually, the link function between the explanatory variables and the variable to be predicted is fixed, even if the parameters are updated over time. For example, in an erasure prediction model, due to a lack of time, it is not always possible to have access to the latest observations. The delay of the variables is thus fixed in a "pessimistic" way to be sure of having the data just necessary for the forecast.

These variables are generally chosen so as to ensure that their current value is available at the moment when the supplier wishes to make the consumption prediction. The setting of this model is preferably pessimistic, as indicated, because it is generally taken a safety margin. When the link function is set, data that can only be intermittently available (from time to time), but not always, is ignored. However, this data may be more recent than parameter data retained and, therefore, provide more relevant information than the parameter data retained. In addition, the provider is never immune to a computer problem, particularly for the retrieval of the chosen parameter data, which could prevent the model from working effectively.

The present invention improves the situation.

It proposes for this purpose a method implemented by computer means, management of an electric power distribution network, the method comprising an estimate of energy demand by consumers, the demand estimation comprising the steps:

a) providing at least one model based on a link function between at least one input parameter of the model and at least one output parameter of the model, b) receiving at the input of a computing device at least one value of the parameter d entrance, and c) applying the model to the received value of the input parameter to output an output parameter value relating to a power demand prediction.

In particular in the present invention, the link function chosen for the model applied in step c) is adaptive and depends on the parameter values received in step b). Steps a) and b) include in particular the operations:

al) providing a plurality of link functions, each specific to a type of model and to at least one input parameter,

b1) receiving at the input of the computing device a plurality of respective input parameter values,

b2) determining, according to a chosen criterion, a ranking in order of relevance of the parameters whose values are received in step b1), and

a2) constructing a model based on at least one link function involving at least one of the most relevant parameters according to the classification of the operation b2), for the subsequent implementation of step c).

Thus, the link function used is not fixed, but it is adaptive according to the most promising parameters for the prediction to be made. In one embodiment, the aforementioned criterion has at least one component relating to the novelty of a value received for one parameter, relative to other values received for other parameters.

Thus, "promising" parameters as indicated above are preferably those whose values are the most recent.

The above-mentioned chosen criterion may be unique or, in a more sophisticated variant, result from a combination of several criteria.

Thus, as a variant or in combination, the criterion chosen may also take into account a planned date for estimating the energy demand. In one embodiment, the model constructed in operation a2) is based on an average of link functions (and thus results from an average of models) each involving at least one parameter among the most relevant parameters according to the ranking of operation b2).

Preferably, this average is a weighted average according to the order of relevance of the parameters determined in the operation b2).

In one embodiment, the ranking in order of relevance of the parameters is done by calculating more precisely the relevance of the models using these parameters. In this embodiment, the operation b2) then comprises:

the construction of a plurality of models using input parameters for which values have been received,

for each model constructed, an estimate of error with respect to a history of values of the input parameters intervening in each model, and

- the ranking of the parameters in order of relevance, according to the estimated error on the models using these parameters.

In this embodiment, the error for each model is preferably estimated by comparing:

the application of the model to previous values of a history of parameter values to anticipate a current value of an output parameter, to

a current value of the output parameter, actually received by the computing device.

In one embodiment, the models of the plurality of models are constructed by learning in a received parameter value history database.

The method may then comprise a target erection step d) of targeted consumers, in case of prediction of a peak demand in step c). The invention also relates to a computer program (the general algorithm of which may correspond to FIG. 4) and comprising instructions for the implementation of the above method, when this program is executed by a processor. The present invention also relates to a computing device (reference SI of Figure 1 discussed below) for managing an electric power distribution network, and implementing an estimate of energy demand by consumers. This device comprises in particular a processing circuit including a processor (reference PROC of Figure 1) for the implementation of the method above.

Other advantages and characteristics of the invention will become apparent on reading the description of embodiment examples below, and on examining the appended drawings, in which:

FIG. 1 illustrates a global system comprising the computing device SI for the implementation of the present invention;

FIG. 2 illustrates more precisely the operation of the computing device SI,

FIG. 3 illustrates the adaptation of the data in history for a particular update,

FIG. 4 illustrates the main steps of a method in a particular embodiment of the invention.

With reference to FIG. 1, the computing device SI (for "information system") is presented in the form of a computer typically comprising a computer processing circuit including:

an input interface IN, to receive historical data to be stored, and to process to establish models by self-learning,

a PROC processor connected to the input interface IN,

a memory MEM2 for storing these data in a database BD history, a working memory MEM cooperating with the processor PROC for the processing of these data, this memory being able to store further instructions of a computer program within the meaning of the invention,

an output interface OUT, for example to send an erasure instruction calculated by the computing device SI, as a function of the processing of the received data.

Of course, an operator can control the operation of the SI device and has in particular an SCR display screen for this purpose. The input data IN can be received via an NW network. It can be typically:

- consumption data (power consumed for example) HRD, which are time stamped as will be seen later,

- meteorological data, with an associated date of the forecast meteorological event,

erase data DEFF for certain voluntary consumers, at times chosen by these consumers for example.

The output data OUT can be sent via the NW network and can correspond to:

erase instructions CEFF intended for identified counters, which can be calculated by the computing device SI, according to the expected demand peaks and the erasing capacities indicated by data DEFF,

- useful information for the energy producer, to assist him in the decision to produce energy from one source (PV photovoltaic, for example) rather than from another (nuclear N, for example), according to the forecasts of consumption needs that can be calculated by the computing device SI,

- etc. In an exemplary embodiment described below, only a history of reliable data of the phenomenon to be predicted (the erasure for example, or simply a peak of request) is available. According to an approach proposed in one embodiment of the invention, a self-learning solution is set up so that when new data arrives, the model relearns the actual parameters of the model. In fact, the data used for forecasting often come from individual meters (for example, from companies). An imponderable thing about this approach is that the data recovery time remains random. However, it appeared that the more recent data used in the model, the better the forecast. Thus, a self-learning solution where the parameters of the model are regularly updated with the arrival of new data is proposed here. The link function that is the basis of the definition of the model is not definitively fixed, but it is adaptive according to the most recent data received.

Not setting the link function and adapting it to the most recent data always makes use of the most up-to-date and therefore most relevant information available. Thus, the invention makes it possible to always produce the best prediction from the available data.

In addition, in the event that a computer failure reaches the level of the computer system (or the SI device presented above) and that the data necessary for the realization of the prediction are not available, it is sufficient to restart the prediction module, which then adapts the link function to the data present in the history database. Here again, the invention makes it possible to always make a prediction even in the case where data recovery is not possible.

Thus, it is proposed to change the parameters with the arrival of new data. In addition, in an example of realization proposed hereafter, the forecast is not simply realized using a single model, but using several models where the final forecast is a weighted average of the together model forecasts, the weight of a model being all the more important as its latest forecasts are better. In one application example, it seeks to predict the erasure of consumers having a consumption contract greater than 250 kW and adhering to an offer with a tariff incentive to erase on a general peak day.

Only a limited history of data is available to make the prediction possible. It is then implemented a computer processing allowing models to relearn the parameters with the arrival of new data (self-learning) and to adapt the link function according to the most recent data available.

With reference to FIG. 2, the data recorded in individual counters 1 to n are read remotely by the supplier and then stored in a database BD. The data is then corrected (for obvious errors) and aggregated. A file of the available data history is then generated. This file includes the polling date, the aggregated power of the portfolio, and the deletion indication.

This file, with the desired start and end dates, is provided to a self-learning module. The self-learning module adjusts the link function of each prediction model according to the data, learns the parameters and makes the prediction on dates passed as parameters.

The prediction is written by the module in an output file which is then read by the computing device SI. The prediction can then be transmitted to an erasure optimizer for example (not shown).

Referring now to Figure 3 to more specifically describe the operation of the self-learning module. The latter begins by reading the explanatory data. First, it calculates the difference in time instants between the data to be predicted and the most recent data. For example, if November 14, 2016 is to be predicted and the last available data deemed relevant is November 10, then an offset is calculated as follows: Delta = 4 days * nb instants per day

The module therefore builds one or more models where the most recent variable ("delayed power" here) is delayed by Delta. It should be noted that it is possible to provide several delayed variables as explanatory variables of the model, possibly different from a delayed power.

Thus, with reference now to FIG. 4, the first step is the calculation of Delta in step S1. Then, in order to learn the parameters of these link functions, the module generates a learning database from which the parameters of the N models are learned (step S2). A prediction is then made from each model for N models from 1 to N (step S3). For a first prediction (T4 test), the applied prediction is simply an arithmetic mean of all the predictions of the N models (step S5). On the other hand, if several predictions have already been made, a weighted average based on the error made by the N models on the last predictions is rather calculated (step S6).

For example, according to the first relevance criterion presented above and relating to the novelty of the data received, a consumption model (and the associated link function) using yesterday's meteorological data would have less weight than a trend model. consumption and using consumption data collected from consumers at a pace of 20 minutes for example. On the other hand, the receipt of more recent meteorological data could reverse this weighting. In addition, according to the second criterion presented above, if a prediction has already been made for the type of day on which the prediction must be made (for example a Monday, or a Tuesday, ...) and a type of instant ( for example at 8h, or 8h30, ...) (T7 test of Figure 4), the final prediction is:

the weighted average of the predictions of the N models (step S8), or the prediction of one of the N models if this model is more efficient for the type of time of day and type of day than the prediction by weighted average (step S9).

On the other hand, if no prediction has been previously calculated for this type of time stamp (instant and day) then the weighted average prediction is used (step S 8).

Of course, the present invention is not limited to the embodiments presented above and admits variants.

For example, the very nature of the models that can be used has many variations and can be complex. The present description does not have to detail such models. It should be remembered that the choice of the model (and the link function between input and output parameters associated with it) is adaptive according to the data received (their novelty, their relevance to a prediction date to achieve, or other).

Furthermore, an order of relevance of the parameters, calculated from the evaluated error for each model using these parameters, has been defined above. This approach nevertheless admits of variants and it is possible for example, in a less sophisticated variant, to define a parameter as directly relevant as a function of the time of reception of the data of this parameter.

In addition, an average of the weighted model results has been described above. It should be noted that the application of a simple arithmetic mean has already made it possible to improve (by 15% according to the tests carried out) the results of applying a single model from a fixed link function to sense of the prior art.

Claims

A method implemented by computer means for managing an electrical energy distribution network, the method comprising an estimate of energy demand by consumers, the demand estimation comprising the steps of:

a) providing at least one model based on a link function between at least one input parameter of the model and at least one output parameter of the model, b) receiving at the input of a computing device at least one value of the parameter d entrance, and

c) applying the model to the received value of the input parameter to output an output parameter value relating to a power demand prediction,

characterized in that the link function chosen for the model applied in step c) is adaptive and dependent on the parameter values received in step b),

and in that steps a) and b) comprise the operations:

2. Method according to claim 1, characterized in that the selected criterion comprises at least one component relating to the novelty of a value received for a parameter, relative to other values received for other parameters.

3. Method according to one of the preceding claims, characterized in that the selected criterion takes into account a date provided for estimating the energy demand.

4. Method according to one of the preceding claims, characterized in that the model constructed in operation a2) is based on an average of link functions each involving at least one parameter among the most relevant parameters according to the classification of operation b2).

5. Method according to claim 4, characterized in that said average is weighted according to the order of relevance of the parameters determined in the operation b2).

6. Method according to one of the preceding claims, characterized in that the operation b2) comprises:

7. Method according to claim 6, characterized in that the error for each model is estimated by comparing:

8. Method according to one of claims 6 and 7, characterized in that the models of the plurality of models are constructed by learning in a database of history of received parameter values.

9. Method according to one of the preceding claims, characterized in that it further comprises a step d) target erase target consumers, in case of prediction of a peak demand in step c) .

10. Computer program characterized in that it comprises instructions for the implementation of the method according to one of claims 1 to 9, when the program is executed by a processor.

11. A computer device for managing an electrical energy distribution network, implementing an estimate of energy demand by consumers, characterized in that it comprises a processing circuit including a processor for the implementation of the process according to one of claims 1 to 9.