Determining an activation over time of a given piece of equipment within a set of pieces of equipment based on collected data
A convolutional neural network-based method for NILM accurately determines equipment activation times using overall consumption data, overcoming data collection challenges and improving adaptability and precision in NILM systems.
Patent Information
- Application Number
- PCT/EP2026/062782
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-06
- Filing Date
- 2026-05-04
- Publication Date
- 2026-11-12
AI Technical Summary
Existing non-intrusive load monitoring (NILM) methods struggle to accurately determine the activation over time of electrical equipment within a set of equipment due to the need for extensive data collection and training on individualized consumption curves, which is costly and socially challenging, leading to statistical biases and imprecise predictions.
A method using a set of convolutional neural networks that transform overall electrical consumption data into time series, apply convolutional blocks with residual units, and calculate activation probabilities without requiring individualized consumption data, allowing for the detection of equipment activation based on a predefined threshold and class activation maps.
Enables accurate determination of equipment activation times with reduced data collection efforts, adapting easily to new types of equipment, and outperforms state-of-the-art models in precision and accuracy, particularly for diverse electrical devices like electric vehicle charging stations.
Smart Images

Figure EP2026062782_12112026_PF_FP_ABST
Abstract
Description
Determining the activation over time of a given piece of equipment within a set of equipment, based on collected data. TECHNICAL FIELD
[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 of disaggregating the load curve of a fleet of equipment in order to determine individual consumptions and to detect periods of activation of the equipment.
[0003] It thus enables the optimization of electricity consumption for users (premises managers) and of electricity production for the supplier or distributor.
[0004] These premises, whether domestic or commercial, 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 building, one can typically find electrical equipment such as a connected television, lighting systems, household appliances, connected sockets, water heaters, electric vehicle charging stations, electrical panels, inverters, etc.
[0006] However, optimizing electricity consumption, or more generally energy consumption, may become increasingly important for better management of the available power on the electrical grid, as well as for economic and environmental reasons. Such optimization can only be achieved through understanding the individual electrical behavior of each piece of equipment.
[0007] Generally, there is no way to measure the electricity consumption of each piece of equipment. Such an approach would require installing an electricity meter for each device, which is undesirable due to the additional cost. Furthermore, such a solution would likely be unacceptable 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 piece of 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 to estimate 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] In their study, they compared the performance of three deep learning architectures: a Denoising Autoencoder (DAE), a recurrent network (LSTM), and a start / stop time regressor. These three models demonstrated superior performance compared to the classical machine learning methods then in use. Furthermore, this work sparked renewed interest in the field within the community, and numerous solutions based on deep learning were subsequently proposed in the literature.
[0014] Approaches based on convolutional neural networks have emerged, such as the one described in Zhang, Chaoyun et al. “Seguence-to-point learning with neural networks for nonintrusive load monitoring” in AAAI Conference on Artificial Intelligence (2016).
[0015] Approaches based on recurrent neural networks (RNNs for "Recurrent Neural Network" in English) have also been proposed.
[0016] Unlike traditional neural networks, RNNs were developed to retain previous information in "memory" while processing current data. The key characteristic of RNNs is the use of recurrent loops, which allow information to be propagated from one time point T to the next. Theoretically, this helps maintain previous states in memory, making RNNs particularly well-suited for sequence prediction tasks, such as machine translation, text generation, or load curve disaggregation. However, basic RNNs are difficult to train in practice, primarily due to the vanishing gradient problem that arises when training on long sequences.
[0017] To address this, variants of RNNs, such as Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Units (GRU), have been developed. These architectures incorporate specific mechanisms to better manage long-term dependencies and mitigate the gradient explosion problem.
[0018] More recently, a new approach has emerged based on "transformer" technology. This type of algorithm was introduced in the article by Vaswani, Ashish et al., "Attention is All you Need," in Neural Information Processing Systems (2017).
[0019] An example of such an algorithm, “BERT4NILM”, is described 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-lntrusive Load Monitoring (NILM'20), Association forComputing Machinery, New York, NY, USA, 89-93 https: / / doi.org / 10.1145 / 3427771.3429390.
[0020] This mechanism allows for 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.
[0021] In general, these proposals suffer from the need to have a learning set associating time series, or sequences of time series, corresponding to global load curves, with labels representing the "truth", that is to say the individualized consumption of the equipment present and connected to the delivery point (PDL).
[0022] These labels must have a granularity corresponding to the resolution of the individualized load curves that we are trying to obtain in prediction. Therefore, if we want curves with 30-minute steps, it is necessary that the models be trained with training sets having a step of no more than 30 minutes.
[0023] Obtaining load curves for the various pieces of equipment in a building (home or business) represents a significant data collection cost, as well as technical and social challenges. Indeed, installing devices to acquire these load curves can encounter difficulties in gaining acceptance from households. Technical difficulties can also arise due to outdated installations.
[0024] The inability to collect load curves for each appliance across all households introduces a statistical bias into the training set. This bias can negatively impact the quality of predictions made by the trained models.
[0025] It emerges that the proposals of the state of the art encounter numerous difficulties. DESCRIPTION OF THE INVENTION
[0026] The invention aims to improve the state of the art. In particular, the problem that the present proposal seeks to solve is to obtain the determination of the activation over time of a given piece of equipment, within a set of equipment located in a given room, based on an overall electrical consumption for this set of equipment.
[0027] To achieve this, a method is proposed for determining the activation over time of a given piece of equipment, within a set of equipment located in a specific room, comprising transforming a stream of measured values of the overall electrical consumption of said set, provided by a measuring device associated with said specific room, into a time series, and dividing said time series into a set of sequences, then, for at least one sequence, the following steps: providing said sequence to a set of convolutional networks associated with said given equipment, including submitting to respective subnets of said convolutional networks to obtain a set of feature maps for each subnet, and then grouping each feature map of each subnet to obtain a representative value of said feature map; for each convolutional network, submission of said vector to a neural layer to determine a probability of activation of said given equipment during the duration corresponding to said sequence7 then calculation of an overall probability of activation from said activation probabilities; if said overall probability of activation is greater than a predefined threshold, for each convolutional network, determination of a class activation map from said set of feature maps and synaptic weights at output of said grouping, then determination of an overall activation map from the activation maps determined for each convolutional network, said overall activation map being representative of said activation over time of said given equipment.
[0028] The process is based on a set of convolutional neural networks, whose learning does not require individualized consumption data for the equipment.
[0029] This process therefore makes it possible to overcome the difficulties of constructing a learning assembly as previously described for prior art proposals.
[0030] Furthermore, it is noted that this process allows for better adaptation to the introduction of a new use or type of equipment. A concrete example is the relatively recent emergence of personal electric vehicle charging stations.
[0031] To enable their detection, the current state of the art would require a new data collection campaign specifically for this equipment, in order to retrain the neural models. The proposed method simply requires indicating the existence of this new equipment for a set of overall consumption curves obtained from a meter installed for the PDL (Point de Livraison - Delivery Point). Implementation is therefore significantly simpler.
[0032] In general, the process only requires for training purposes that the person in charge of the premises (home, company, etc.) indicates the types or classes of equipment available to them.
[0033] It's clear that retrieving this data is far less expensive and complicated than collecting complete labels, such as consumption curves per device. Based on a 30-minute time step, such an approach requires collecting tens of thousands of values per room, compared to just one data point per device (approximately ten per room) using the proposed method.
[0034] 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 activation over time of said given equipment is determined as a function of the product of said global activation card and said at least one sequence; said set of convolutional networks comprises networks that differ by at least one meta-parameter, for example by kernel size for a convolutional block constituting said respective subnetwork; said subnetwork comprises a succession of residual blocks generating an increasing number of feature maps, each residual block comprising a succession of convolution blocks; During a training phase, a superset of convolutional neural networks is trained on a training set. Then, a score is determined for each network on a validation set, and the set of convolutional neural networks is selected based on these scores. During a training phase, the set of convolutional neural networks is trained on a training set associating sequences with an indicator of the presence or absence of the given equipment. A power consumption curve for the given equipment is provided, based on its activation over time and an estimated power consumption for the given equipment.
[0035] Another object concerns a method for characterizing electrical consumption, including the implementation of the method for determining the activation over time of a given piece of equipment as previously described, for a plurality of pieces of equipment.
[0036] Another object relates to a system comprising at least one processor and memory adapted to implement a process as previously described.
[0037] Another aspect of the invention relates to a computer program, suitable for implementation on an electronic system, comprising code instructions which, when executed by a processor of said electronic system, carry out the steps of the process described above. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Other aspects, objectives, advantages, and features of the invention will become clearer upon reading the following detailed description of preferred embodiments thereof, given by way of non-limiting example, and made with reference to the accompanying drawings in which: Figure 1 illustrates a context of use of the process and device according to embodiments; Figure 2 is an illustrative flowchart of a process according to one embodiment; Figure 3 illustrates a software architecture associated with the process according to one embodiment; Figure 4 illustrates a possible architecture for a subnetwork of a convolutional neural network, according to one embodiment; Figure 5 illustrates a possible architecture for a residual unit, according to one embodiment. DETAILED DESCRIPTION OF SPECIFIC IMPLEMENTATION METHODS
[0039] Figure 1 shows a set of equipment, Ei, E2, E3, ..., E n , located in a specific location L.
[0040] The defined area L can correspond to a geographically limited space within which a given user can place and control connected devices. This space can correspond to a personal dwelling (apartment, house, etc.), or to the premises of a company, a store, etc.
[0041] The determined premises may possibly have extensions outside of a limited space, when for example certain connected objects are located remotely: an electric car charging station may be located in a garage, a garden, or even open onto the road, and be outside the main space corresponding to the dwelling, while still being within the user's control perimeter.
[0042] Equipment can be of different kinds, or types: they are said to correspond to different uses. Generally speaking, it refers to any equipment connected to an electrical power distribution network and consuming this energy, whether continuously or intermittently.
[0043] Examples include: household appliances (ovens, refrigerators, hot plates, etc.), lighting systems, heating and thermal regulation systems, security systems (motion detectors, surveillance cameras, etc.) etc.
[0044] A device C is associated with the specified room L, and is adapted to provide measurement data of an overall load, or consumption, for all the equipment associated with the specified room L. This device can measure the energy consumed on the internal NE distribution network within the specified room L. The measured values can typically represent apparent power.
[0045] This device is typically a so-called "smart" electricity meter, that is, one adapted to provide such data, usually in the form of a stream of numerical values.
[0046] This could include, for example, a Linky™ type electricity meter in France, or a reading device to be connected to the main circuit breaker of premises L, such as those marketed by the company Ecojoko.
[0047] A suitable GTW gateway can also be provided to communicate with the measurement device C, in order to acquire a stream of measurement values of overall electrical consumption provided by the measurement device C.
[0048] The GTW gateway may include data processing capabilities. It may also have interfaces with the measurement device C and a NET telecommunications network.
[0049] The interface with the measurement device C can be a wireless interface, in order to minimize wired connections within room L and for ergonomic reasons. To achieve this, a wireless module may need to be connected to the measurement device C to ensure wireless connectivity between it and the GTW gateway. This module could be a local radio transmitter (LRT) using the Zigbee™ protocol, for example.
[0050] According to a particular embodiment, the GTW gateway may include a Raspberry Pi type microprocessor, RAM memory (between 4 GB and 8 GB, for example), and mass storage, for example of the eMMC type, of about 16 GB.
[0051] The interface to the NET telecommunications network can be wired (Ethernet™) or wireless (Wi-Fi...). More specifically, 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 (the Internet). The NET telecommunications network can therefore be viewed as a set of subnets.
[0052] The NET telecommunications network allows the GTW gateway to communicate with an S service platform.
[0053] Different role sharing arrangements between the GTW gateway and the S services platform can be considered.
[0054] The S service platform can be in charge of managing the GTW gateway, including software updates, fault management, etc.
[0055] Data and processing can also be transmitted from the GTW gateway to the S service platform. This data and processing may correspond to tasks that cannot be performed locally, particularly because the computing power of an S service platform is required. The S service platform can be deployed on a server farm or, abstracted as a cloud computing server, its computing power can be adapted according to the task to be performed, as submitted by the GTW gateways.
[0056] These treatments can correspond to training predictive models on training sets.
[0057] Indeed, training is costly in terms of computing power and execution memory. It is therefore advantageously executed on a service platform (S) which can be adequately sized for this type of processing, whereas the GTW gateway can only have limited resources since it is a device designed to be deployed at the user's premises: it therefore meets strict cost and sizing constraints.
[0058] The GTW gateway can advantageously implement the inference process allowing the extraction of the electrical consumption of an individual piece of equipment within a set of equipment.
[0059] In such an implementation, the model(s) trained by the service platform S can be uploaded to the GTW gateway.
[0060] Other arrangements and distributions of processing between a local GTW gateway (possibly subdivided into several separate devices) and a remote S service platform can also be implemented.
[0061] In particular, in inference, part of the processing can be moved to the S service platform, in particular to aggregate data from different locations (in which case the data can be anonymized beforehand).
[0062] Figure 2 illustrates a flowchart representing a division into steps of the treatments of the proposed process.
[0063] In various embodiments, this process can be implemented by an electronic system comprising the GTW gateway and the S service platform. Different choices can be made regarding the division of tasks between the gateway and the service platform. These choices may be dictated by the optimization of data transmissions between the gateway and the service platform, the computing and data manipulation capabilities of the GTW gateway, the gateway's autonomy, and service continuity in the event of a communication failure with the S service platform.
[0064] This division into steps may be seen as arbitrary and solely intended to make the description clearer. A person skilled in the art may indeed choose to proceed with alternative divisions of the treatments, in particular by proposing subdivisions of certain steps, or by grouping others together.
[0065] This processing is applied, or implemented, to a flow of measurements of overall electrical consumption provided by the measurement device C. This overall electrical consumption therefore corresponds to the set of equipment, Ei, E2, E3, ..., E n , located in the designated room L associated with this measuring device C.
[0066] Generally, the measurements are average power measurements for a given time step. This time step can be 30 minutes.
[0067] This flow of measurements, corresponding to a load curve, can correspond to a variable duration. It can represent historical data or a current flow for near real-time processing.
[0068] One of the goals of the proposed method is to analyze the energy consumption of a room L, particularly with a view to potential optimization. Therefore, it is important to have sufficient historical data to capture different behaviors and identify patterns and recurring events. A history of approximately three days is recommended.
[0069] This measurement stream is transformed into a time series. This time series can represent a time window in which each value represents a measurement value (for example, an average power) for an interval within that time window. For example, we could have a time interval, or time step, of 15 seconds and a time series of 480 values, which represents a total time window of 480 x 15 = 2 hours.
[0070] Within a time series, the value assigned to a time interval, or not, can correspond to a resampling of the values in the measurement stream. For example, if more than one value in the measurement stream is available for the same time interval, a preprocessing step can establish an average value (for example) in order to construct the time series.
[0071] In an SI step, the time series is broken down into a set of sequences.
[0072] As previously discussed, the time series can correspond to a load curve extending over several hours, or even several days. For subsequent processing, this large volume of data can be subdivided into sequences corresponding to shorter time periods. Preferably, the sequences correspond to constant durations.
[0073] We can denote by x this sequence which associates a value x(t) with a time t. This time t is a discrete time, corresponding to the measurement or resampling step.
[0074] From this sequence, we seek to determine the activation over time of a given piece of equipment among the set of equipment in a room L, that is to say whose electrical consumption contributes to the sequence x(t).
[0075] Activation over time can be understood as determining whether the equipment in question is activated or not at each instant t (that is, whether the equipment is on or off, or, more precisely, whether it is consuming electrical energy or not). This allows us to determine the equipment's on and off times.
[0076] From such a method of determining the activation of a piece of equipment, one can characterize an electrical consumption by implementing this method for a plurality of equipment, or even for all the known equipment in room L.
[0077] Figure 3 illustrates a software architecture enabling this determination of the activation of a piece of equipment.
[0078] A set of N convolutional neural networks (10i, 10j, 10N) is planned for each device whose activation over time is to be detected. In other words, if the activation of several devices is to be detected, several sets of convolutional neural networks must be planned, each specialized for a specific device through a learning phase.
[0079] These convolutional networks are trained on a training set during a learning or training phase prior to using the networks in prediction to determine equipment activations.
[0080] This learning package is specific to the equipment in question.
[0081] The training set consists of associations between sequences x(t) from time series and labels indicating the presence or absence of the given equipment.
[0082] For example, for a given piece of equipment (e.g., a refrigerator), the training set consists of time series sequences containing a contribution from that equipment, as well as sequences not containing that equipment. These sequences are associated with an indicator of the equipment's presence or absence. In this way, neural networks can be trained to recognize whether or not the equipment contributes to a given time series sequence x(t).
[0083] The problem targeted by the learning is therefore a binary classification assigning a binary label, with a first class corresponding to the detection of the equipment and a second class corresponding to the non-detection of the equipment.
[0084] As mentioned previously, it is worth noting that the learning set is particularly easy to create since it only requires: From electricity consumption curves provided by a C measurement device (such as a smart electricity meter), A declaration of possession or non-possession of a given electrical equipment by a manager of the associated premises L.
[0085] It is not necessary to have a specific load curve for this equipment during the learning phase.
[0086] According to one embodiment, during the training phase, a superset of convolutional networks is trained on a training set, then a score is determined for each convolutional network on a validation set, and the set of convolutional networks is selected from said scores.
[0087] For example, we can train 15 convolutional models, for example on significantly different training sets, then based on the scores obtained on validation sets, we select the best ones, for example the 5 best, which we then use for the prediction, or exploitation, phase.
[0088] The training and validation sets can form a partition of the set of available associations. In particular, for each convolutional model, a different partition can be defined (for example, randomly), which generates a diversity of training and therefore of trained models.
[0089] Typically, a training set representing 80% of the available associations is chosen, and a validation set representing the remaining 20%.
[0090] These convolutional networks can also offer other elements of diversity.
[0091] They may, for example, differ by one or more meta-parameters, such as convolution kernel size.
[0092] In particular, it is proposed to vary the size of the kernel of the first convolutional block of the residual units (or "resunit") of the convolutional networks: for example 5, 7, 9, 15 and 25 to form 5 different networks.
[0093] Three networks can be trained for each kernel size, resulting in a total of 15 networks, from which the 5 best can be selected, for example. The architecture of the possible convolutional networks will be described in more detail later.
[0094] Having several different convolutional networks, especially with different kernel sizes, allows for multiple levels of explainability.
[0095] In a step S2, the sequence x(t) is provided to this set of N convolutional networks, previously trained as previously indicated, associated with a studied equipment.
[0096] Each 10i, 10j, 10N convolutional network is preferentially structurally identical to the others, and they differ only in their training and / or meta-parameters as previously explained.
[0097] For clarity of the figure, only a convolutional network, 10j, is detailed.
[0098] Providing the sequence to a convolutional network includes submitting it to a subnetwork 11 of that convolutional network 10j.
[0099] Figure 4 illustrates a possible architecture for this subnetwork, in which it comprises a succession of, for example, 3 residual units, 110i, IIO2, IIO3.
[0100] Figure 5 illustrates a possible architecture for a 110j residual unit.
[0101] This can include a succession of, for example, 3 convolutional blocks ll1, III2, IH3.
[0102] These convolutional blocks may themselves include a convolution layer, a batch normalization layer, and an activation layer.
[0103] This type of convolutional network was initially introduced in the article by He, K., Zhang, X., Ren, S., & Sun, J. (2016), “Deep residual learning for image recognition” In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).
[0104] ResNet (short for Residual Network) is a convolutional neural network architecture that marked a major turning point in the field of deep learning. ResNet was designed to solve a fundamental problem in training deep networks: the performance degradation that occurs as the network depth increases. Contrary to what one might think, adding more layers to a network does not guarantee better accuracy. In fact, beyond a certain point, deep networks become more difficult to optimize, and their performance can even deteriorate. To overcome this obstacle, ResNet introduces an ingenious mechanism called a residual connection (or "skip connection"), which allows layers to transmit their outputs directly to layers further down the network, bypassing one or more intermediate layers.These connections facilitate backpropagation of the gradient, reduce problems related to gradient vanishing, and allow for the efficient training of networks with up to several hundred, or even thousands, of layers.
[0105] According to one embodiment, the network comprises 3 residual units that differ in the depth (or number of channels) of the output feature map. For example, there may be 64 channels at the output of the first residual unit 110i, 128 channels at the output of the second residual unit IIO2, and another 128 channels at the output of the third residual unit IIO3.
[0106] Furthermore, the convolutional blocks Illi, III2, III3 constituting each residual unit 110i (with i=l, 2,3) can differ in the size of the kernels of the convolution layer.
[0107] For example, the kernel size for the first convolutional block llli can differ from one convolutional network to another, as previously explained. Kernels of sizes 5, 7, 9, 15, or 25 are possible.
[0108] The kernel sizes for the second and third convolutional blocks III2, III3 can be fixed from one convolutional network to another, for example 5 and 3 respectively.
[0109] The convolutional blocks III, III2, and III3 each include an ID convolutional layer that acts as a sliding filter on the input sequence to extract local patterns or regularities over time t, thus capturing local dependencies within a sequence. It transforms the input data into a new set of more abstract representations called feature maps.
[0110] Batch normalization (or "BatchNorm") or layer normalization (or "LayerNorm," according to common English terminology) can then be applied to stabilize and accelerate network training. This operation involves normalizing the activations of the previous layer by subtracting the mean and dividing them by the standard deviation calculated on the mini-batch. This reduces sensitivity to scaling variations and improves gradient propagation.
[0111] The choice between batch or layered normalization may depend on the volume of data to be processed. Batch normalization can be used with a reasonably sized load curve, but layered normalization should be chosen when this load curve becomes too large.
[0112] Layer normalization was introduced by the article by BA, Jimmy Lei, KIROS, Jamie Ryan, and HINTON, Geoffrey E., "Layer normalization". arXiv preprint arXiv:1607.06450, 2016.
[0113] Finally, a ReLU (for "Rectified Linear Unit") activation function introduces essential non-linearity into the model.
[0114] Furthermore, the sequence x(t) is reintroduced as input to each residual unit 110i (and therefore to the first convolutional block llli).
[0115] The output of this subnetwork 11 yields a set of feature maps fk(t), where k is an index of the feature map. If we have 128 channels at the output of the last residual unit, III3, we have ke[l, 128],
[0116] Then, for this subnetwork 11, the feature maps fk(t) are grouped by a neural layer 12, to obtain a representative fk value per feature map.
[0117] This neural layer 12 is adapted to perform grouping or subsampling, or "pooling" in English.
[0118] This step aims to reduce the size of the feature maps to fk(t). Specifically, the reduction is performed on the temporal dimension. It is also adapted to obtain a single, time-independent fk value for each temporal map. This value can be an average across the entire feature map or a GAP (global average pooling) value.
[0119] We therefore obtain, at the output of this layer 12, a vector (fk) of dimension the number of feature cards (for example 128).
[0120] In an S3 step, for each convolutional network 10i, the vector (fk) is submitted to a neural layer 13 adapted to determine an activation probability Pi of the equipment studied during the duration corresponding to the sequence x(t).
[0121] This neuronal layer 13 is typically a linear layer, also called a fully connected layer or dense layer.
[0122] The grouping network 12 can include a set of neurons, each neuron corresponding to a feature map fk(t) provided as input and, therefore, to a value fk representative of this map.
[0123] These neurons can be fully connected to linear layer 13, by connections associated with synaptic weights Wk.
[0124] Linear layer 13 outputs an activation probability that is a binary value. Linear layer 13 can consist of a single neuron, whose value is determined by the vector fk and the synaptic weights Wk linking it to each of the neurons in pooling layer 12.
[0125] We can then write the following for the probability Pi of the "equipment activation" class for a 10i convolutional network: Pi = * [1^ ) \ k /
[0126] A bias can also be expected in the neuron 13 activation equation.
[0127] The function s() is an activation function. It could, for example, be the sigmoid function.
[0128] This prediction Pi corresponds to the expected behavior due to the learning of the convolutional network from association between sequences x(t) and indicators of presence or absence of the equipment.
[0129] In a step S4, a module 20 is planned to calculate an overall probability P of activation of the equipment studied from the activation probabilities Pi determined by each convolutional network 10j.
[0130] This overall probability can be an average over the N activation probabilities obtained from the N convolutional networks.
[0131] We can then write: N i=l
[0132] In step S5, this overall probability P can be used to determine whether the equipment under study is active or not during the duration of the sequence x(t). This information can be useful, but one can also seek to determine the activation over time of the equipment under study, that is, as previously explained, detect the times t of activation and deactivation of this equipment.
[0133] For example, this probability can be compared to a threshold to determine whether the equipment is active or not in the sequence x(t). A binary threshold around the pivot value 0.5 can be used.
[0134] If P < 0.5, the equipment under study is considered not to be active. The process can be interrupted and another sequence x(t) can then be considered (in Figure 2, looping back to step S2).
[0135] If P > 0.5, then we consider that the equipment studied is active for the duration corresponding to the sequence x(t).
[0136] In this case, in an S6 step, an adapted module 30 determines a class activation map, CAM (for "Class Activation Map" in English) from the set of feature maps fk(t) and synaptic weights Wk at the output of the grouping step (or layer) 12.
[0137] A class activation map (CAM) is a technique used to identify which parts of a signal (here, a sequence x(t)) a convolutional neural network uses to make a prediction, particularly in classification. The idea is that by weighting each feature map with the weights of the target class, we can visualize which regions contribute most to the prediction. Here, we are interested in the "equipment activation" class.
[0138] We can then write, for each 10j convolutional network: CAM t (0 Wk ' fk(t)
[0139] The CAMi(t) values of each card can be normalized, for example by dividing them by the maximum value of the card. _ z CAM t (t) CAM ' m (maxCCWQ)
[0140] This yields N normalized class activation maps, whose values are necessarily between 0 and 1.
[0141] In an S7 step, a suitable module 40 can determine a global activation map CAM(t) from the N activation maps CAMi(t) (normalized or not).
[0142] This global CAM(t) activation map can be written as: N CAM^ -^ CÀM^t) i=l
[0143] This global activation map CAM(t) is representative of the activation over time of the equipment studied.
[0144] However, according to one embodiment, better accuracy can be obtained for the activation over time of the given equipment by performing, in a step S8, a product of this global activation map with the processed sequence x(t).
[0145] Indeed, simply using the average of the activation maps across the entire set does not achieve the same level of performance: the resulting "mask" is often imprecise and points to numerous periods before or after the actual device activations. Therefore, utilizing the shape of the aggregate input load curve, x(t), allows for greater accuracy by targeting exclusively the peaks responsible for the activation of the device in question.
[0146] The product can be a point-by-point product, meaning that each value in the global activation map CAM(t) is multiplied point by point by the processed sequence x(t) of the time series. In other words, we obtain a product vector p(t) such that: Vt, p(t) = CAM(t) xx(t)
[0147] A sigmoid function is then applied to obtain values between 0 and 1, because multiplying the global activation map and the input curve (expressed in kWh) does not allow us to directly obtain activation probabilities.
[0148] Experimental studies conducted by the inventors have shown that omitting step S8 results in a decrease in accuracy of over 68%. However, depending on the circumstances, it may still be useful to determine activations over time based solely on the overall activation map (i.e., without implementing step S8).
[0149] In one embodiment, an electrical consumption curve for the equipment under study is determined from its activation over time and an estimated consumption. This estimated consumption can be a nominal consumption provided by the person in charge of room L or based on a type of equipment. For example, an estimated consumption can be arbitrarily assigned for a "dishwasher" type appliance, or the person in charge of the room can provide the consumption as indicated on the technical data sheet for the equipment they have.
[0150] The proposed method has been tested experimentally on various real-world disaggregation cases for the purpose of estimating the quantity of energy and detecting activation periods, as well as on publicly available test sets (or "benchmarks").
[0151] Several metrics exist to allow the comparison of the performance of disaggregation models on different applications (i.e., correctly allocated amount of energy or correct activation detection).
[0152] For each measurement, T represents the total number of intervals (number of time steps or number of days, weeks, or months) and y t represents the actual consumption value and y t the consumption value predicted by the proposed method.
[0153] We can then define classic regression metrics.
[0154] We can first define "classic" regression metrics MAE (for "Mean Average Error" in English, or mean absolute error) and RMSE (for "Root Mean Squared Error" in English, or mean squared error): T MAE = ^£ Wt ~ yt ] t=i i T RMSE = ^\yt -yù 2 t=i
[0155] We can also define the precision P (or "accuracy" in English) from the rates of true positives (TP), true negatives (TN), false negatives (FN) and false positives (FP) by: TP + TN P ~ TP + TN + FP + FN
[0156] An Fl score can also be defined as follows: TP ri _ 2 x _ 2TP + FP + FN
[0157] The MR (Matching Ratio) is defined as the ratio of the overall accuracy of the match between the actual and estimated energy consumption for a piece of equipment. It can be defined as follows: = 2 1 min (y t ,y t ) iXimax (y t ,y t )
[0158] These metrics can be used to compare the proposed process with state-of-the-art solutions, and in particular with the CRNN proposal, which is the only proposal that can be deployed with indicators of the presence or absence of equipment as labels.
[0159] A CRNN (for "Convolutional Recurrent Neural Network") is a hybrid architecture that combines the advantages of convolutional neural networks (CNNs) and recurrent neural networks (RNNs).
[0160] The table below gives performance metric values for different consumption curve bases and different equipment for the proposed process and for a CRNN network.
[0161] [Table 1] CRNN Proposed Method Fl MAE RMSE MR Fl MAE RMSE MR Dishwasher 0.54 44.8 242.28 0.2 0 50.5 295.5 0 (1) Kettle 0.695 10.8 125.6 0.48 0.3 353.6 664.2 0.1 Microwave 0.16 4.3 64.64 0.08 0.0 5.6 72.3 0.0 Washing machine 0.137 19.6 172.73 0.025 0.0 19.5 176.3 0.0 Dishwasher 0.464 40.4 273.8 0.032 0.0 35.5 252.4 0.0 (2) Kettle 0.763 20.9 191.72 0.295 0.3 61.9 267.6 0.1 Microwave 0.126 6.9 81.3 0.0 0.1 19.8 77.06 0.027 Dishwasher 0.324 10.6 116.65 0.11 0.0 10.8 125.9 0.0 (3) Shower 0.892 9.7 131.8 0.795 0.7 39.8 489.34 0.5 Washing machine 0.04 14.7 151.06 0.008 0.0 14.6 151.8 0.0 (4) Vehicle 0.742 230.0 850.17 0.464 0.3 2371 2819 0.1 electric Average 0.376 38.5 227.166 0.226 0.16 273.6 552.43 0.065
[0162] References (1), (2), (3), (4) represent 4 banks of electrical consumption curves.
[0163] References (1) to (3) represent 3 public banks, respectively REFIT, UKDALE and IDEAL.
[0164] These three databases present consumption values at fine time steps (5-10 seconds) for overall load curves and household appliances. UK-DALE offers data from 5 English homes collected from 2012 to 2015, and REFIT from 23 English homes collected from 2015 to 2017. IDEAL is more recent and is described in the article by Pullinger M, Kilgour J, Goddard N, Berliner N, Webb L, Dzikovska M, Lovell H, Mann J, Sutton C, Webb J, Zhong M. “The IDEAL household energy dataset, electricity, gas, contextual sensor data and survey data for 255 UK homes”. Sci Data. 2021 May 28;8(1):146. doi: 10.1038 / s41597-021-00921-y. PMID: 34050194; PMCID: PMC8163809.
[0165] It can be noted that the data collected in these studies were collected at "high frequency" time steps (one point collected per second) and the instrumented devices happen to be "small" devices consuming little energy, such as the washing machine, the dishwasher or the kettle.
[0166] The applicants also compiled a database (reference 4 in the table above) containing data from 24 households, recording the household's overall energy consumption and the electric vehicle's charging curve. A vehicle charging station corresponds to significantly higher electricity consumption and is therefore of a very different nature from that of the equipment studied in studies (1), (2), and (3).
[0167] The table above shows that the proposed process outperforms the CRNN model for almost all metrics and equipment. The average line highlights the superior performance achieved for the four metrics used.
[0168] Furthermore, it has been experimentally demonstrated that the performance of the proposed process is equivalent to other state-of-the-art proposals based on learning from individualized consumption curves for each piece of equipment, although it uses approximately 20 to 500 fewer labels collected for this learning.
[0169] The proposed method therefore makes it possible to determine the activation over time of a given piece of equipment within a set of equipment located in a specific room, with a low constraint on the acquisition of training data. This allows for the rapid creation of a new training set, particularly if a new piece of equipment or a new use case needs to be considered. New equipment can thus be taken into account (for example, air conditioning in a region where it was previously uncommon, household robots, etc.).
[0170] This information about the activation of a given piece of equipment can be used in various ways.
[0171] As previously discussed, multiple, or even all, of the equipment in a given room can be analyzed, allowing for the characterization of that room. It was also previously shown that each piece of equipment could be associated with a typical (e.g., average) power consumption in order to deduce an estimated consumption based on information about equipment activation.
[0172] Extracting the power consumption of an individual piece of equipment within a group of devices allows for various practical applications. As mentioned previously, consumption curves can be extracted for multiple pieces of equipment, thus enabling the characterization of a room's power consumption.
[0173] This extraction (or characterization) for a local L is important for several reasons, and can be exploited in different ways.
[0174] From a general point of view, this extraction is of great interest.
[0175] It helps, in particular, to manage energy consumption from the perspective of the consumer, the energy supplier, and the environment as a whole. This last point addresses current and future requirements regarding the environmental impact of energy consumption by users (both residential and industrial).
[0176] Furthermore, in order to design smart buildings, more efficient distribution systems, and so on, it is important to have more precise and numerous metrics. Such a mechanism for extracting electricity consumption data from a home or business can also contribute to building a knowledge base that facilitates research and development in these areas.
[0177] Extracting and / or characterizing electricity consumption can be used to optimize the energy production of the electricity supplier.
[0178] The characteristics of a large number of end consumers can be transmitted, in anonymized form, to a concentrator that can compile consumption statistics. This concentrator can be an application module on a service platform accessible to the energy supplier.
[0179] As an example, it is possible to establish an average consumption, or activation, curve per type of equipment.
[0180] Based on these statistics, the energy supplier can adapt its production network. In particular, it can build a reliable forecast of customer usage and thus anticipate increases in demand at its production sites or, conversely, decreases in demand, thereby enabling better management of the electricity grid.
[0181] For example, it can characterize a peak load related to electric vehicle charging stations at the end of the day, and search for mechanisms to respond to it.
[0182] The characterization process can therefore enable energy suppliers to optimize energy production, in particular to avoid waste through overproduction, thanks to the transmission of characterization data to a service platform accessible to the energy supplier.
[0183] Also, an optimization module can be included to allow for the optimization of the Ei, E2, E3, ..., E equipment fleet n located in a specific location L based on consumption data obtained from predictive models.
[0184] According to one embodiment, an optimization module can determine control orders for one or more pieces of equipment among those connected to the network in order to respond to internal or external constraints of the specified room L, or of the building to which the specified room L belongs.
[0185] In particular, a possible constraint may be to balance consumption, or at least to constrain fluctuations in consumption beyond a given threshold.
[0186] Thus, in the event of a consumption peak (detectable from the consumption curve provided by measurement device C), characterizing the electrical consumption across the equipment fleet can help determine a probable cause of this peak (either through consumption data or, failing that, through activation data). It is then possible to define control commands to smooth out electrical consumption.
[0187] For example, some equipment can be temporarily switched off or forced into a less power-consuming state, for a duration corresponding to the activity of the equipment that was the source of the peak consumption.
[0188] As an illustrative example, a kitchen heating system can be lowered when an oven is switched on: its operating time is limited so that lowering the heating is not very detrimental to the inhabitants.
[0189] In general, heating, water heating, and electric vehicle charging, in particular, are equipment that can be controlled to coincide with off-peak hours of local consumption and off-peak hours of electricity production by the supplier.
[0190] Furthermore, characterizing consumption makes it possible to identify the share of energy consumption per piece of equipment and thus provide an energy cost for each. This allows users to be offered specific actions to reduce energy consumption per piece of equipment and to receive a precise return on investment indication, since it is possible to know the energy and financial contribution associated with each piece of equipment: for example, it is possible to indicate potential savings on their energy bill by lowering the heating by one degree.
[0191] Similarly, the characterization of electrical equipment over a given time window can be one of the inputs for a load control model based on energy predictions that take this characterization as input. For example, this control model could send commands to turn electrical devices on or off based on past characterization data.
[0192] Furthermore, the optimization module can provide the user with a representative overview of consumption and activation data for all equipment. This data can be aggregated and consolidated to provide a comprehensive view of their electricity usage.
[0193] The optimization module can detect optimization opportunities and, possibly after validation by a user from local L, transmit appropriate command orders.
[0194] For example, it can detect standby equipment and electricity consumers: plugged-in household appliances, for instance. It can also identify heating systems used at night. In some cases, it may be beneficial to limit these uses, and the optimization module can suggest control commands aimed at reducing this consumption without impacting the comfort of the occupants (or at least with an energy savings / comfort loss ratio greater than one).
[0195] Of course, other optimization scenarios can be implemented by the optimization module, taking advantage of the detailed knowledge of the electrical uses of the premises obtained by the characterization process described.
[0196] Various other use cases are also described in the scientific or commercial literature related to non-intrusive load characterization (NILM).
[0197] It is clear that the characterization obtained allows for the optimization of electricity consumption by consumers and the optimization of electricity production by electricity suppliers.
[0198] Of course, the present invention is not limited to the examples and embodiment described and illustrated, but is defined by the claims. In particular, it is susceptible of numerous variations accessible to those skilled in the art.
Claims
DEMANDS 1. Method for determining the activation over time of a given piece of equipment, within a set of equipment (E1, E2, E3, E1). n ) located in a specific location (L), comprising a transformation of a flow of measured values of the overall electrical consumption of said set, provided by a measuring device (C) associated with said specific location (L), into a time series, and a segmentation (SI) of said time series into a set of sequences (x(t)), then, for at least one sequence, the steps of: supply (S2) of said sequence to a set of convolutional networks (10j) associated with said given equipment, including submission to respective subnets (11) of said convolutional networks to obtain a set of feature maps (fk(t)) for each subnet, and then grouping each feature map of each subnet to obtain a value (fk) representative of said feature map; for each convolutional network, submission (S3) of said vector to a neuronal layer (13) to determine an activation probability (Pi) of said given equipment during the duration corresponding to said sequence7 then calculation (S4) of an overall activation probability (P) from said activation probabilities; if said overall activation probability is greater than a predefined threshold, for each convolutional network (10i), determination (S6) of a class activation map (CAM) from said set of feature maps (fk(t)) and the synaptic weights at the output of said grouping, then determination (S7) of an overall activation map from the activation maps determined for each convolutional network, said overall activation map being representative of said activation over time of said given equipment.2.A method according to claim 1, wherein said activation over time of said given equipment is determined (S8) as a function of the product of said global activation card and said at least one sequence.
3. A method according to any one of claims 1 or 2, wherein said set of convolutional networks comprises networks that differ by at least one meta-parameter, for example by a kernel size for a convolutional block constituting said respective subnetwork (11).
4. A method according to any one of claims 1 to 3, wherein said subnetwork (11) comprises a succession of residual blocks generating an increasing number of feature maps, each residual block comprising a succession of convolution blocks.
5. A method according to any one of the preceding claims, wherein, during a training phase, a superset of convolutional networks is trained on a training set, then a score is determined for each convolutional network on a validation set, and said set of convolutional networks is selected from said scores.
6. A method according to any one of the preceding claims, wherein, during a learning phase, said set of convolutional networks is trained on a training set associating sequences with an indicator of the presence or absence of said given equipment.
7. A method according to any one of the preceding claims, wherein a power consumption curve of said given equipment is provided from said activation over time of said given equipment and an estimated consumption for said given equipment.
8. A method for characterizing power consumption comprising an implementation of the method for determining an activation over time of a given piece of equipment according to any one of the preceding claims, for a plurality of equipment.
9. System comprising at least one processor and memory adapted to implement a method according to one of the preceding claims.
10. Computer program capable of being implemented on an electronic system, the program comprising code instructions which, when executed by a processor of said electronic system, carry out the steps defined in any one of claims 1 to 8.
