Characterization for the optimization of the electrical consumption of a set of equipment connected to an electrical network

Abstract

Characterization for the optimization of the electrical consumption of a set of equipment (E1, E2, E3, …, En) located in a room (L), comprising the transformation of a flow of measurements of an overall electrical consumption, provided by a measuring device (C) into a time series, which is provided as input to at least one first predictive model (MP1) adapted to provide consumption data for an equipment and at least one second predictive model (MP2) adapted to provide activation data for an equipment, in which the second predictive model (MP2) comprises a first sub-model (MP1') corresponding to the first predictive model and a second sub-model (CC) adapted to determine an activation probability (P) from the output of the first sub-model. Figure for the abstract: Fig. 1

Description

Title of the invention: Characterization for the optimization of the electrical consumption of a set of equipment connected to an electrical network FIELD OF THE INVENTION

[0001] The present invention relates to the non-intrusive determination of the electrical consumption of electrical equipment located in a specific room.

[0002] The invention relates more particularly to a method for disaggregating the load curve of a fleet of equipment in order to determine individual consumptions and to detect the periods of activation of the equipment.

[0003] It thus allows the optimization of electricity consumption for users (premises managers) and electricity production for the supplier or distributor.

[0004] These premises, whether domestic or business, contain a growing number of electrical devices whose nature and behavior are extremely diverse, both in terms of consumption and activation rate per period of time.

[0005] In a residential premises, one can typically find electrical equipment such as a connected television, lighting systems, household appliances, connected sockets, water heaters, charging stations, electrical panels, inverters, etc.

[0006] However, it may be increasingly important to optimize electrical consumption, or more generally energy consumption, for economic and ecological reasons. Such optimization can only be achieved through knowledge of the electrical behaviors individualized by equipment.

[0007] In general, there is no measurement of the electricity consumption of each piece of equipment. Such an approach would require the installation of an electricity meter for each piece of equipment, which is not desirable due to the additional cost that this may generate. Furthermore, such a solution would probably not be acceptable to users.

[0008] Non-intrusive load monitoring (or NILM) approaches have been proposed.

[0009] Load curve disaggregation aims to estimate the individual energy consumption of each appliance, or equipment, and / or the activation of the on / off state, using only the aggregated total load curve (apparent power consumed for a dwelling). This aggregated total load curve is available via a smart electricity meter.

[0010] Load curve disaggregation was initially approached as a linear combination problem. Early research investigated the use of combinatorial optimization algorithms with the objective of estimating the proportion of energy consumption at each time step and for each active device. Subsequently, hidden Markov models were the preferred approach for several years.

[0011] In recent years, proposals have been made based on neural networks and deep learning.

[0012] One of the first publications relating to this family of approaches is the article by Kelly, J., & Knottenbelt, W. “Neural nilm: Deep neural networks applied to energy disaggregation”, in Proceedings of the 2nd ACM international conference on embedded Systems for energy-efficient built environments (pp. 55-64), 2015.

[0013] Other approaches are based on convolutional neural networks, such as the one described in Zhang, Chaoyun et al. “Sequence-to-point learning with neural networks for nonintrusive load monitoring” in AAAI Conference on Artificial Intelligence (2016).

[0014] More recently, a new approach has emerged based on the technology of "transformers". This type of algorithm was introduced in the article by Vaswani, Ashish et al. “Attention is AU you Need.” In Neural Information Processing Systems (2017).

[0015] An example of such an algorithm “BERT4NILM” is described in the article by Zhenrui Yue, Camilo Requena Witzig, Daniel Jorde, and Hans-Amo Jacobsen “BERT4NILM: A Bidirectional Transformer Model for Non-Intrusive Load Monitoring” in Proceedings of the 5th International Workshop on Non-Intrusive Load Monitoring (NILM'20). Association for Computing Machinery, New York, NY, USA, 89-93 https: / / doi.org / 10.1145 / 3427771.3429390.

[0016] This mechanism allows a partial disaggregation of the load curve: the individual load (or consumption) of certain types of equipment can be determined, but this proves impossible, or too imprecise for other types of equipment.

[0017] It has also been proposed to focus on determining the activity of a piece of equipment (i.e. whether it is switched on or not) over a unit of time, rather than on the disaggregation of its load curve.

[0018] There is therefore a need to improve the current state-of-the-art proposals. Summary of the invention

[0019] The invention aims to combine the approaches by disaggregation of the load curve and by determination of the activation of equipment. In particular, it aims to integrate both approaches optimally, that is, by minimizing the resources required.

[0020] For these purposes, according to a first aspect, the present invention can be implemented by a method for characterizing the electrical consumption of a set of equipment located in a given room, comprising the transformation of a flow of measurement values ​​of an overall electrical consumption of said set, provided by a measuring device associated with said given room, into a time series, said time series being provided as input to at least one first predictive model adapted to provide consumption data for at least one respective equipment among a first subset of said set and at least one second predictive model adapted to provide activation data for at least one respective equipment among a second subset of said set,wherein said at least one second predictive model comprises a first sub-model corresponding to said at least one first predictive model and a second sub-model adapted to determine an activation probability from the output of said first sub-model, and wherein said at least one predictive model is trained on a predetermined training set and said at least one second predictive model is trained from said first predictive model, by transfer, then from a second training set.

[0021] According to preferred embodiments, the invention comprises one or more of the following features which can be used separately or in partial combination with each other or in total combination with each other: - said at least one first predictive model is a neural network comprising an embedding module, a transformer module and a multi-layer perceptron module; - said second sub-model comprises two fully connected dense layers; - control commands intended for said equipment are determined by an optimization module based on said consumption data and said activation data. These control commands can then be transmitted to the equipment by means of interfaces and protocol adaptations or to devices adapted to equipment control; - said consumption and activation data are transmitted, in an anonymized form, to a concentrator of a second service platform, adapted to establish statistics on the consumption of a set of specific premises, in order to adapt the electricity production of an electricity supplier; - a training module, comprising said at least one first and second predictive model, is stored in a secure structure within a gateway, and in which a training phase comprises the transmission of said module to a service platform through a telecommunications network, the execution of said training module by said service platform, in order to train said predictive modules, then the transmission of said predictive models to said gateway; - a learning classification module is adapted to select said learning sets according to equipment present in said specific room: - when training said second predictive model from said second training set, only said second sub-model is adapted; - when training said second predictive model from said second training set both said first sub-model and said second sub-model are adapted.

[0022] Another aspect relates to a computer program comprising instructions for implementing a method as previously described, when said method is implemented on an information processing platform.

[0023] Another aspect relates to a gateway comprising a processor adapted to implement a method as previously described, optionally in collaboration with a service platform through a telecommunications network.

[0024] Other characteristics and advantages of the invention will appear on reading the following description of a preferred embodiment of the invention, given by way of example and with reference to the appended drawings. BRIEF DESCRIPTION OF THE FIGURES

[0025] The accompanying drawings illustrate the invention: [Fig.l] schematically illustrates a context of use of a proposed method.

[0026] [Fig.2] schematically represents a functional view of a system comprising a gateway in collaboration with a service platform, according to an embodiment of the invention.

[0027] [Fig.3] shows a diagram of a functional view of a gateway and the use of predictive models for characterizing electrical consumption, according to one embodiment.

[0028] [Fig.4] illustrates an implementation of an architecture of a predictive model based on the “BERT4NILM” network, according to one embodiment.

[0029] [Fig.5] illustrates an example of functional architecture for a predictive model of a second type, according to one embodiment.

[0030] [Fig.6] illustrates an embodiment of a learning phase according to one embodiment.

[0031] DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION

[0032] In [Fig.l] are represented a set of equipment, Eb E2, E3, ..., En, located in a specific room L.

[0033] The determined premises L may correspond to a geographically limited space within the perimeter of which a given user can have the connected objects and control them. This space may correspond to a personal home (apartment, house, etc.), or to the premises of a company, a store, etc.

[0034] The determined premises may possibly have extensions outside of a limited space, when for example certain connected objects are remote: a charging station for an electric car may be located in a garage, a garden, or even overlook the roadway, and be outside the main space corresponding to the accommodation, while being within the user's control perimeter.

[0035] The equipment can be of different types. Generally speaking, it is any equipment connected to an electrical energy distribution network NE and consuming this energy continuously or not.

[0036] For example, we can cite: - household appliances (ovens, refrigerators, hobs- heating...), - lighting systems, - heating and thermal regulation systems, - security systems (motion detectors, surveillance cameras, etc.

[0037] A device C is associated with the determined premises L, and is adapted to provide measurement data of a global load, or consumption, for all the equipment associated with the determined premises L. This device can measure the energy consumed on the distribution network NE inside the determined premises L. The measured values ​​can typically represent an apparent power.

[0038] This device is typically a so-called "smart" electric meter, i.e. one adapted to provide such data, generally in the form of a stream of digital values. This may in particular be a Linky™ type electric meter in France, for example.

[0039] A GTW gateway may also be provided, adapted to communicate with the measuring device C, in order to acquire a flow of measurement values ​​of an overall electrical consumption provided by the measuring device C.

[0040] The GTW gateway may include data processing means. It may also have interfaces with the measuring device C and a telecommunications network N.

[0041] The interface with the measuring device C may be a radio interface, in order to minimize the wired connectivity within the room L and for ergonomic reasons. To do this, a radio module may have to be connected to the measuring device C in order to ensure radio connectivity between it and the GTW gateway. This module may be a local radio transmitter (ERL) using the Zigbee™ protocol, for example.

[0042] According to a particular embodiment, the GTW gateway may comprise a Raspberry Pi type microprocessor, RAM memory (between 4 GB and 8 GB, for example), and a mass memory, for example of the eMMC type, of approximately 16 GB.

[0043] The interface to the telecommunications network N can be wired (Ethernet) or wireless (Wi-Fi, etc.). More precisely, this telecommunications network can be composed of a local network to which a telecommunications gateway is connected in order to allow access to the public telecommunications network (Internet). The telecommunications network N can therefore be seen as a set of subnetworks.

[0044] The telecommunications network N allows the GTW gateway to communicate with a service platform S.

[0045] Different role sharing between the GTW gateway and the S service platform can be envisaged.

[0046] The service platform S may be in charge of managing the GTW gateway, including software updates, fault management, etc.

[0047] Data and processing may also be transmitted from the GTW gateway to the service platform S. These processings and data may correspond to tasks that cannot be performed locally, in particular because the computing power of a service platform S is necessary. The service platform S may be deployed on a server farm or abstracted in the form of a cloud computing server. Its computing power may be adapted according to the task to be performed, submitted by the GTW gateways.

[0048] These treatments may correspond to the training of predictive models on learning sets.

[0049] According to one embodiment, as illustrated by the example of [Fig.2], a secure digital structure SEC is provided within the GTW gateway. The security is provided to prevent any third party not having the appropriate keys from being able to read the content of the digital structure.

[0050] According to one embodiment, the training module can be stored in this secure SEC structure within said gateway.

[0051] This training module may comprise a set of predictive models, MP, typically multi-layer artificial neural networks, and computer code, MI, making it possible to implement the training steps.

[0052] During a training phase, this training module can be transmitted to the service platform S via the telecommunications network, in a step SI.

[0053] In a step S2, the training module is executed by the service platform S. The transmitted predictive models MP are thus trained on the basis of training sets (described later).

[0054] In a step S3, the training module, or only the predictive models MP, are transmitted to the GTW gateway.

[0055] The training step S2 is expensive in terms of computing power and execution memory. It is thus advantageously executed on a service platform S which can be adequately sized for this type of processing, while the GTW platform can only have limited resources since it is a device intended to be deployed at users' premises: it therefore meets strong cost and sizing constraints.

[0056] It should however be noted that only the training phase is implemented on the service platform S. As will be seen later, the use of the predictive models MP in order to characterize the overall consumption of the dwelling L can be carried out entirely locally by the GTW gateway.

[0057] In the case where access to the telecommunications network N is interrupted, the GTW gateway can continue to operate for prediction. Training can also be implemented, but since the resources of the GTW gateway are limited, the time required for the convergence of the training will necessarily be longer.

[0058] The method therefore allows optimized mixed collaboration between cloud computing processing and local processing (or edge computing).

[0059] This training module can be included in a "container", which is a data structure containing computer code and elements dependent on this code, adapted to be executed on a virtual machine. Thus, among other advantages, the code can be executed independently of the operating system and the hardware infrastructure of the equipment on which it is to be executed: the same code can thus be executed on the GTW gateway or on the service platform S.

[0060] These containers, stored in the secure digital structure SEC, also make it possible to protect the predictive models MP and the associated computer code MI from possible theft and hacking.

[0061] [Fig.3] illustrates an embodiment of the use of predictive models for characterizing the electrical consumption of all the equipment located in the determined room L.

[0062] As mentioned previously, the GTW gateway can acquire flow of measurement values ​​(or flow of measurements, by misuse of language) of an overall electrical consumption of all equipment connected to the electrical network NE of the premises L.

[0063] The GTW gateway may comprise a pre-processing module MT (for example in the secure digital structure SEC) adapted to transform this flow of measurement values ​​into a time series.

[0064] This time series is constructed to be provided as input to MPI, MP2 predictive models, of the multi-layer neural network type.

[0065] This time series may represent a time window in which each value represents a measurement value for an interval of this time window.

[0066] For example, we can have a time interval, or time step, of 15 seconds, and a time series of 480 values, which represents a total time window of 480x15 = 2 hours.

[0067] Within the time series, the value assigned to a time interval may correspond to a resampling of the values ​​of the measurement stream. For example, if more than one value of the measurement stream is available for the same time interval, the MT preprocessing module may establish an average value (for example) in order to construct the output time series.

[0068] Each predictive model is adapted for a particular type of equipment. Indeed, each type of equipment has its own electrical behavior. As will be seen later, each predictive model is specifically trained on a training set corresponding to a type or class of equipment, so that at the end of this training phase, the model is adapted for this type of equipment.

[0069] These predictive models can be divided into two subsets: - a first subset of predictive models, MPI, is adapted to provide consumption data for equipment of a first subset of equipment (among the set of equipment); - a second subset of predictive models, MP2, is adapted to provide activation data for the equipment of a second subset of equipment (among the set of equipment).

[0070] In other words, depending on their type, the equipment connected to the NE electricity network can be classified into two categories, depending on whether disaggregated electricity consumption can be determined from the overall consumption acquired. via the measuring device C, or if we can only determine their activation, that is to say if they are on or off, for each time step considered.

[0071] This distribution in a first class, corresponding to the first subset of predictive models, MPI, or in a second class, corresponding to the second subset of (second) predictive models, MP2, can be done according to a profile of the individual load curve of the equipment.

[0072] Such a classification has for example been proposed in the article by George W. Hart, “Nonintrusive Appliance Load Monitoring”, in Proceedings of the IEEE, December 1992, pp. 1870-1891. The proposed classification is as follows: - permanent consumption uses: this category includes uses that consume constant power, 24 hours a day, 7 days a week. These include, for example, fire or security alarm devices, telecommunications devices such as modems, internet access gateways, etc. - On / Off uses: this category consists of appliances with only two states. An OFF state (or off) where they do not consume power and an ON state (or on) where they consume approximately constant power. Many common appliances in the residential sector belong to this category. Examples include kettles, light bulbs, refrigerators, freezers, toasters, coffee machines, microwave ovens, "traditional" ovens, certain heating systems, etc. - Finite-state uses: this category includes uses that go through several distinct states where they consume a constant power for each state. This is typically the case for equipment following an automatic cycle formed by a succession of states. Examples include washing machines, dryers, dishwashers, etc. - Infinite state uses: this category includes equipment whose power consumption can vary continuously over a certain range of values, generally depending on a user setting. Examples include variable power vacuum cleaners, etc.

[0073] It is also possible to define a finer, or different, classification than these categories.

[0074] For example, in the case of heating systems, considered above in the ON / OFF usage category, it may be easier to extract the load curve of a heat pump than that of an electric convector. We can therefore consider that each of them belongs to a different class.

[0075] Depending on the type of equipment, a first or a second subset of predictive models (respectively first, MPI, and second, MP2, predictive models) are used.

[0076] This choice of directing a type of equipment towards one or other of the subsets of predictive models can be made according to a parameter setting, which can be set by the supplier of the GTW platform and / or the MPI, MP2 predictive models.

[0077] It may depend on the ability of the predictive models of the first subset to provide consumption data for this given type of equipment from the overall consumption data (disaggregation). This ability may be evaluated by the designer of the predictive model in question, and this evaluation may make it possible to assign this predictive model to this type of equipment, or, if the evaluation is not considered sufficiently positive, to a predictive model of the second subset.

[0078] This setting may change over time. For example, new predictive models may be made available (for example due to new training). It is possible that new predictive models of the first subset offer better capabilities to disaggregate global consumption curves. In this case, it may be chosen to change the assignment of a type of equipment from the second subset to the first subset of predictive models. This change may be made by an update of the preprocessing module MT, for example by downloading a computer code or configuration parameters from the service platform S.

[0079] According to one embodiment, the preprocessing module MT can be provided to define a first subset of equipment and a second subset of equipment, the two subsets being disjoint and forming a partition of the total set of equipment.

[0080] The MT preprocessing module can provide the time series of a piece of equipment as input to a predictive model corresponding to this piece of equipment. This predictive module belongs - to the first subset of (first) MPI predictive models if the equipment belongs to the first subset of equipment, or, - to the second subset of (second) MP2 predictive models, if the equipment belongs to the second subset of equipment.

[0081] According to one embodiment, the MT preprocessing module may comprise a table associating with each piece of equipment or each type of equipment a predictive model in one or other of the subsets.

[0082] According to one embodiment, the disaggregation of the load curve is implemented only for ON / OFF uses. It is also possible, as indicated previously, to distinguish only a sub-part of this equipment for the disaggregation of the load curve.

[0083] In other words, according to this embodiment, the preprocessing module MT can be provided to (by means of a preconfigured table, for example) provide the received time series to a predictive module of the first subset of predictive models MPI if the equipment belongs to certain types of ON / OFF usage equipment, and to a predictive module of the second subset MP2 in other cases.

[0084] As mentioned previously, the MPI, MP2 predictive models can be artificial neural networks.

[0085] According to one embodiment, the predictive models of the same subset, MPI, MP2, can be structurally identical. They differ by their internal state (synaptic weights of the different constituent layers) which are fixed during the learning phase: as will be seen later, each predictive model is the subject of learning on a specific learning set.

[0086] According to one embodiment, a neural network comprising an immersion module, a transformer module and a multi-layer perceptron module.

[0087] [Fig.4] illustrates an implementation of such an architecture based on the “BERT4NILM” network. This type of network was introduced in the article by Zhenrui Yue, Camilo Requena Witzig, Daniel Jorde, and Hans-Arno Jacobsen “BERT4NILM: A Bidirectional Transformer Model for Non-Intrusive Load Monitoring” in Proceedings of the 5th International Workshop on Non-Intrusive Load Monitoring (NILM'20). Association for Computing Machinery, New York, NY, USA, 89-93. https: / / doi.org / 10.1145 / 3427771.3429390

[0088] A BERT4NILM architecture is based on the well-known BERT architecture described in the paper by Devlin, Jacob, Chang, Ming-Wei, Lee, Kenton, et al. “Bert: Pre-training of deep bidirectional transformers for language understanding”, arXiv preprint arXiv: 1810.04805, 2018.

[0089] A BERT4NILM network consists of an embedding module M1, a transformer module M2 and a multilayer perceptron module M3, chained sequentially.

[0090] The embedding module Ml (or "embeddings" in English) comprises an extraction of the characteristics (or "features" in English) of the input Ll, which increases the dimensionality of the data. A "pooling" operation (or "pooling") is then applied in a layer L3.

[0091] The output of this layer L3 is then added to a position embedding matrix L4, in the resultant, L5, can form the input of the transformers (or "transformers" in English) L6, L7 ..., L1 1, of the transformer module M2.

[0092] The transformer module is formed of several layers, each layer having several attention heads. This concept of attention in an artificial neural network was introduced by the article by Vaswani, Ashish, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. “Attention is All you Need.” in Neural Information Processing Systems (2017).

[0093] The multilayer perceptron module M3 comprises a deconvolution layer L12, and two linear layers L13, L14, in order to provide an output vector of the same dimension as the input.

[0094] The input to the MPI predictive model is a time series obtained from a stream of measurement values ​​of overall electrical consumption.

[0095] This time series is representative of the flow of measurement values ​​of the overall electrical consumption of all the equipment connected to the electrical distribution network. The output L14 of the predictive model corresponding to a given equipment is also a time series, but this is representative of a flow of measurement values ​​of the predicted (or estimated) electrical consumption for this given equipment.

[0096] In other words, the MPI predictive model makes it possible to extract a consumption curve for a given piece of equipment from the consumption curve for all the equipment.

[0097] This consumption data provided by the MPI predictive model can be of the same dimension as the input vector L1, i.e. represent a time window of the same length with an identical step.

[0098] By aggregating the consumption data over time, a consumption curve can be obtained for a given piece of equipment. The start and end of this curve correspond to the start and stop of the prediction process that has just been described. According to one embodiment, preferably, this process is implemented continuously, or in any case over long periods, in order to be able to determine results that can be used by the MO optimization module.

[0099] The same time series, obtained from the measurement stream provided by the measuring device C, can be provided to a plurality of predictive models of the first set of predictive models, MPI. Each being trained for a particular type of equipment, will make it possible to determine consumption data for the equipment belonging to this type. In other words, a plurality of load curves are obtained as output.

[0100] Furthermore, the preprocessing module MT can be provided to (by means of a preconfigured table, for example) provide the received time series to a predictive module of the second subset MP2, in the cases previously described.

[0101] These second MP2 predictive models are structurally different from the first MPI models since they aim to perform a classification and not a regression.

[0102] [Fig.5] illustrates an example of functional architecture for an MP2 predictive model of the second subset.

[0103] This second predictive model MP2 comprises a first sub-model MPI' corresponding to the first predictive model MPI and a second sub-model CC adapted to determine an activation probability P from the output of the first sub-model, MPI'.

[0104] The first MPI' sub-model may be structurally identical to the first MPI predictive model: same number of layers, same number of neurons per layer, etc. Thus, both may correspond to the implementation illustrated in [Fig.4]. On the other hand, they may differ in their internal state, determined by the learning phase. In other words, according to embodiments, the synaptic weights may differ between the two MPI, MPI' models.

[0105] The second CC sub-model may comprise two fully connected dense layers. They aim to summarize the information of the representation obtained at the output of the first MPI' sub-model in order to obtain a classification probability P.

[0106] This classification P represents a probability of activation of a given equipment in the time window corresponding to the input data, i.e. whether this equipment is switched on (i.e. consumes electrical energy) or not. As previously explained, a plurality of predictive models MP2 can be provided, each corresponding to a given type of equipment.

[0107] This probability can be simply thresholded to provide a binary activation, i.e. an estimate of whether the equipment is on or off during the corresponding time period.

[0108] The outputs of the different predictive models, of the two subsets, are transmitted to a characterization module MC responsible for processing its individual results of each predictive model to provide an exploitable result, for example by an optimization module MO.

[0109] This MC characterization module aggregates the outputs of the different predictive modules and can also be provided to manage the temporal aspect by serializing the different outputs as a function of time.

[0110] For example, it can reconstruct a consumption curve from the outputs of the predictive models of the first subset which correspond to time windows. To do this, it can put them end to end.

[0111] It can also construct an activation curve by matching the output of the second subset's predictive models to each corresponding time interval.

[0112] It is thus possible to obtain a series of curves, either of consumption or of activation, corresponding to the different equipment connected to the electrical distribution network.

[0113] Generally speaking, the predictive models MPI, MP2, together make it possible to obtain a characterization of the electrical consumption of a set of equipment Eb E2, E3, ..., En located in a given room L connected to the electrical network, from a flow of values ​​of overall electrical consumption measurements provided by a measuring device C associated with the given room L.

[0114] The characterization of the electrical consumption for a room L is important for different reasons, and can be exploited in different ways.

[0115] From a general point of view, this characterization is of great interest, both from a purely scientific point of view and from a technological point of view.

[0116] In particular, it helps to control energy from the point of view of the consumer, the point of view of the energy supplier and the overall point of view. This last point meets current and future requirements on the ecological impact of users' energy consumption (both residential and industrial).

[0117] Furthermore, in order to design smart buildings, more efficient distribution systems, etc., it is important for the scientific community to have more precise and numerous metrics. Such a mechanism for characterizing the electricity consumption of a household, or a business, can contribute to building a knowledge base to facilitate research and development in these areas.

[0118] The characterization of the electrical consumption can be used for the optimization of the energy production of the electrical energy supplier.

[0119] The characterizations of a large number of end consumers can be transmitted, in an anonymized form, to a concentrator which can establish statistics on consumption. This concentrator can be an application module on a service platform accessible to the energy supplier.

[0120] As an example, it is possible to establish an average consumption or activation curve by type of equipment.

[0121] Based on these statistics, the energy supplier can adapt its production network. In particular, it can build a reliable prediction of customer usage and thus predict increases in the load of its production points or, on the contrary, reductions in loads, thus enabling it to reduce its exposure to the speculative risks of the inter-producer electricity market.

[0122] The characterization process can therefore allow energy suppliers to optimize energy production, in particular to avoid waste by overproduction, thanks to the uploading of characterization data to a service platform accessible to the energy supplier.

[0123] Also, an optimization module MO can be provided in the GTW gateway or in the service platform S, in order to allow the optimization of the fleet of equipment Eb E2, E3, ..., En located in a given location L as a function of the consumption data and the activation data obtained from the predictive models (via the characterization module).

[0124] Preferably, the optimization module MO is embedded (in the form of a software or application module) within the GTW gateway. This allows optimization to be carried out on the basis of locally available information: this minimizes the transmission of data to the Internet, which is advantageous in terms of personal data management and saving bandwidth and network resources. In addition, optimization can continue even if the link to the telecommunications network and the service platform S is cut.

[0125] The optimization module can determine control orders for one or more pieces of equipment among those connected to the network in order to respond to these internal or external constraints to the determined room L, or to the building to which the determined room L belongs.

[0126] In particular, a possible constraint may be to balance consumption, or at least to constrain consumption fluctuations within a given threshold.

[0127] Thus, in the event of a consumption peak (detectable from the consumption curve provided by the measuring device C), the characterization of the electrical consumption on the equipment fleet can make it possible to determine a probable cause of this peak (either by consumption data or, failing that, by activation data). It is then possible to determine control orders to smooth the electrical consumption.

[0128] For example, certain equipment may be momentarily switched off or forced into a less consuming state, for a duration corresponding to an activity of the equipment source of the consumption peak.

[0129] As an illustrative example, a kitchen heating system can be turned down when an oven is switched on: its operating time is limited in time so that lowering the heating is of little disadvantage to the inhabitants.

[0130] Furthermore, the characterization of consumption makes it possible to identify the share of energy consumption per piece of equipment and thus makes it possible to provide an energy cost for each. It is thus possible to propose to the user specific actions to reduce energy consumption per piece of equipment and to provide a precise return on investment time since it is possible to know the energy share and financial information associated with each piece of equipment: it is thus possible to indicate a possible saving on their energy bill by lowering the heating by one degree.

[0131] Also, the characterization of electrical equipment over a given time window can be one of the inputs of a load control model which would be based on energy predictions which would take this characterization as input. For example, this control model could send orders to turn on or off electrical devices according to the past characterization carried out.

[0132] Furthermore, the MO optimization module can provide the user with a representative status of the consumption and activation data of all the equipment. This data can be aggregated and consolidated so as to provide an overall view of its electrical usage.

[0133] The optimization module can detect optimization opportunities and, possibly after validation by a user of the premises L, transmit suitable control orders.

[0134] For example, it can detect equipment on standby and consuming electricity: connected household appliances, for example. It can also detect heating systems used at night. In certain cases, it may be interesting to limit these uses and the MO optimization module can propose control orders aimed at reducing these consumptions which should not impact the comfort of the inhabitants (or in any case present an energy saving / loss of comfort ratio greater than one).

[0135] Obviously, other optimization scenarios can be implemented by the optimization module, MO, by taking advantage of the detailed knowledge of the electrical uses of the premises obtained by the characterization method described.

[0136] Various other use cases are further described in the scientific or commercial literature related to non-intrusive load characterization (NILM).

[0137] It is clear that the characterization obtained by the GTW gateway allows the optimization of the electrical consumption of consumers and the optimization of the electrical production of electrical energy suppliers.

[0138] As indicated, this characterization method is based on predictive models MPI, M2, which have undergone a training phase.

[0139] During this: - the first MPI predictive model(s) of the first set are trained on a predetermined training set, - the second predictive model(s) MP2 of the second set are trained from the first predictive models, MPI, by transfer, then from a second training set.

[0140] [Fig.6] illustrates an embodiment of this learning phase.

[0141] In this embodiment, the training is deployed on the service platform S. As described previously, in one embodiment, a training module can be transmitted from the GTW gateway to the service platform S through the telecommunications network (dotted arrows).

[0142] A database DB contains a predetermined training set for the first MPI predictive models and a second training set for the second MP2 predictive models.

[0143] This learning set associates time series corresponding to flows of values ​​of measurements of an overall electrical consumption to - consumption data, in the case of the first training set (i.e. corresponding to the first predictive models) - activation data, in the case of the second learning set

[0144] These consumption and activation data correspond to labels representative of a reality. The learning sets can be constituted by carrying out experimental studies in the laboratory or within real user premises, by having measuring equipment making it possible to record the consumption and / or activation of individual equipment.

[0145] Public learning databases exist and can be used, such as the REDD, UK-DALE or REFIT databases. These three databases present consumption values ​​at fine time steps (5-10 seconds) for overall load curves and domestic equipment. REDD offers data from 6 American houses collected in 2011, UK-DALE from 5 English houses collected from 2012 to 2015, and REFIT from 23 English houses collected from 2015 to 2017.

[0146] According to one embodiment, a classification module for learning, MCA, is provided for selecting learning sets from the content of the learning sets contained in the database DB.

[0147] The selection can be made according to the equipment contained in the determined premises L. The examples selected in the database to constitute the learning sets correspond to those presenting a strong resemblance with the determined premises L. This resemblance is estimated from the types of the main equipment (which can be provided by a list), but also by other impacting elements: surface area, individual house or apartment, etc.

[0148] Non-electrical equipment can also be taken into account. For example, the presence of a fireplace or a stove can impact electricity consumption for heating-related uses.

[0149] This selection therefore makes it possible to take into account in the learning only examples relevant to the determined location L. We therefore obtain models MPI, MP2 predictors adapted to the determined location L, which allows more relevant predictions.

[0150] A predictive model is trained for each device, i.e. for each label, whether it is formed from consumption data (for a given device) or activation data (also for a given device).

[0151] A predictive MPI model is trained for each of the devices connected to the electrical network of the determined premises L.

[0152] In the case where the first predictive models are of the BERT4NILM type, their training can be carried out in accordance with the article describing this type of architecture, in particular in terms of cost function (or “loss function” in English).

[0153] The second predictive models are trained a first time by transfer, that is to say that the internal state of the first predictive models are copied to form the respective second predictive models. In other words, the knowledge acquired during the first training phase (captured by the state of the first predictive models) is reused, or transferred, to the second predictive models MP2.

[0154] Then, the second training set previously mentioned is used for a second “sub-phase” of training. This second training aims to specialize the second predictive models in the task of determining (or predicting) activation data.

[0155] In other words, it aims to mainly train the second CC sub-network intended to precisely summarize the information produced by the first sub-network into an activation probability P.

[0156] For this training, a BCE cost function (for "Binary Cross Entropy" in English, or "binary cross entropy" in French) can be used between the activation probability P at the output and the label provided by the second training set (as selected by the classification module for MCA training, if applicable). The BCE cost function is a very common function, defined as a measure of the difference between two distribution probabilities.

[0157] This training makes it possible to converge the internal state of the second predictive models to specialize them in a classification task (determination of an activation P). In particular, it seeks to fix the state of the second CC sub-network constituting the second predictive models MP2.

[0158] According to one embodiment, the internal state of the first subnetwork MPI' is fixed during this training and therefore remains the same as that of the respective first corresponding MPI model. In other words, only the second CC sub-model is adapted during training,

[0159] According to another alternative embodiment, the internal state of the first MPI' sub-network is not fixed during this training. In other words, both the first MPI' sub-model and the second CC sub-model are adapted during the training.

[0160] Experimental studies have demonstrated the effectiveness of the proposed approach. In particular, the proposed method for characterizing electrical consumption offers quality measurements (precision, recall rate, etc.) superior to state-of-the-art proposals for a set of varied types of electrical equipment.

[0161] Some state-of-the-art proposals may achieve better results for certain types of equipment, but to the extent that a given room L includes equipment of various types, these specialized mechanisms are not satisfactory.

[0162] The proposed method, on the contrary, makes it possible to achieve high performance in the case of a set of equipment of various types (refrigerator, freezer, heating system, computer, microwave ovens, washing machine, etc.).

[0163] In particular, the proposed method allows high performance in the case of refrigeration equipment (freezers, refrigerators, etc.) to obtain consumption data per equipment. It also allows high performance in obtaining activation data for household appliance type equipment.

[0164] Of course, the present invention is not limited to the examples and the embodiment described and shown, but is defined by the claims. It is in particular susceptible of numerous variants accessible to those skilled in the art.

Claims

Claims

1. Method for characterizing the electrical consumption of a set of equipment (Eb E2, E3, En) located in a given room (L), comprising the transformation of a flow of measurement values ​​of an overall electrical consumption of said set, provided by a measuring device (C) associated with said given room (L), into a time series, said time series being provided as input to at least one first predictive model (MPI) adapted to provide consumption data for at least one respective equipment among a first subset of said set and at least one second predictive model (MP2) adapted to provide activation data for at least one respective equipment among a second subset of said set,wherein said at least one second predictive model (MP2) comprises a first sub-model (MPI') corresponding to said at least one first predictive model and a second sub-model (CC) adapted to determine an activation probability (P) from the output of said first sub-model, and wherein said at least one predictive model (MPI) is trained on a predetermined training set and said at least one second predictive model (MP2) is trained from said first predictive model, by transfer, then from a second training set.,

2. Method according to the preceding claim in which said at least one first predictive model (MPI) is a neural network comprising an embedding module (M1), a transformer module (M2) and a multilayer perceptron module (M3).

3. Method according to the preceding claim, wherein said second sub-model (CC) comprises two fully connected dense layers.

4. Method according to one of the preceding claims in which control orders intended for said equipment are determined by an optimization module (MO) as a function of said consumption data and said activation data.

5. Method according to one of the preceding claims, in which a training module, comprising said at least one first and second predictive model, is stored in a secure structure (SEC) within a gateway (GTW), and in which a phase training comprises the transmission of said module to a service platform (S) through a telecommunications network (N), the execution of said training module by said service platform, in order to train said predictive modules, then the transmission of said predictive models to said gateway (GTW).

6. Method according to one of the preceding claims, in which a learning classification module (MCA) is adapted to select said learning sets according to equipment present in said determined room (L).

7. Method according to one of the preceding claims, wherein when training said second predictive model (MP2) from said second training set, only said second sub-model (CC) is adapted.

8. Method according to one of claims 1 to 6 wherein when training said second predictive model (MP2) from said second training set both said first sub-model (MPI') and said second sub-model (CC) are adapted.

9. Computer program comprising instructions for implementing a method according to one of the preceding claims, when said method is implemented on an information processing platform.

10. Gateway (GTW) comprising a processor adapted to implement a method according to one of claims 1 to 8, optionally in collaboration with a service platform (S) through a telecommunications network (N).

Images