# US 12,224,587 B2 · Extraction of instantaneous renewable generation from net load measurements of power distribution systems > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,224,587 B2 - Document ID: US12224587B2 - Status: Granted - Kind: B2 - Application number: 17/656,465 - Filed: 2022-03-25 - Published: 2023-09-28 - Granted: 2025-02-11 - Earliest claimed: 2022-03-25 - Assignees: Mitsubishi Electric Research Laboratories, Inc. - Applicants: Mitsubishi Electric Research Laboratories, Inc. - Inventors: Hongbo Sun; Ashwin Shirsat; Kyeong Jin Kim; Jianlin Guo - Web page: https://patentgrep.com/patent/US12224587B2 ## CPC - H02J 3/004: {Generation forecast, e.g. methods or systems for forecasting future energy generation} - G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] - G06N 3/045: Combinations of networks - G06N 3/0455: Auto-encoder networks; Encoder-decoder networks - G06N 3/049: Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs - G06N 3/096: Transfer learning - G06Q 50/06: Energy or water supply - H02J 3/003: {Load forecast, e.g. methods or systems for forecasting future load demand} - H02J 3/381: {Dispersed generators} - G06N 3/048: Activation functions - H02J 2101/24: Photovoltaics - H02J 2103/30: Simulating, planning, modelling, reliability check or computer assisted design [CAD] of electric power networks - H02J: ELECTRIC POWER NETWORKS; CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR ## Abstract A computer-implemented method is provided for performing energy disaggregation of a distribution system-level net-load measurements using continuous-point-on-wave (CPOW) measurement units. The method uses a processor coupled with a memory stored instructions implementing the method using neural networks including an encoder network, a feature extractor, a separator network, a decoder network stored in the memory, wherein the instructions, when executed by the processor carry out at steps of the method include generating net-load time series data from voltage and current measurements via the CPOW measurement units, generating a compressed latent space representation from the net-load time series, converting the net-load time series into time-frequency domain, passing time domain cotextual information with the converted time-frequency domain representation of net-load time series to the feature extractor, estimating two weight matrices to be multiplied with an output from the encoder network and learning temporal features of a native load and a photovoltaic (PV) generation, transforming weighted latent representation corresponding the native load and the PV generation into time-domain representations, and predicting the native load and the PV generation at distribution system-level from the transformed time domain representations corresponding to the native load and PV generations. ## Claims ### Claim 1 A computer-implemented method for performing energy disaggregation of a distribution system-level net-load measurements on a distribution system using continuous-point-on-wave (CPOW) measurement units, wherein the method uses a processor coupled with memory stored instructions implementing the method using neural networks including an encoder network, a feature extractor, a separator network, a decoder network stored in the memory, wherein the instructions, when executed by the processor carry out steps of the method, comprising: generating net-load time-series data from voltage and current measurements via the CPOW measurement units; generating, using the encoder network, a compressed latent space representation from the net-load time series; converting the net-load time series into the time-frequency domain using a short-time Fourier transform (STFT) method; passing time-domain cotextual information with the converted time-frequency domain representation of net-load time series to the feature extractor trained to accurately extract harmonic features; estimating, using the separator network trained to assign weight to the compressed latent space representation, two weight matrices to be multiplied with an output from the encoder network and learning temporal features of a native load and a photovoltaic (PV) generation; transforming, using the decoder network, the weighted compressed latent representation corresponding the native load and the PV generation into time-domain representations from outputs of the encoder network and the separator network; and predicting, using a post-disaggregation adjustment method, the native load and the PV generation at distribution system-level from the transformed time-domain representations corresponding to the native load and the PV generation. ### Claim 2 (depends on 1) The method of claim 1, wherein the encoder network performs a one-dimensional (1D) convolution operation to transform the net-load time series into the compressed latent space representation of net-loads. ### Claim 3 (depends on 1) The method of claim 1, wherein the feature extractor network identifies, using the STFT method, the harmonic features introduced in the distribution system by inverter-based PV generation. ### Claim 4 (depends on 1) The method of claim 1, wherein the separator network uses a temporal convolutional network (TCN) to learn the temporal features, wherein the TCN includes sequential layers of stacked dilated 1D TCN blocks, wherein each of the stacked dilated 1D TCN blocks consists of a gated residual network. ### Claim 5 (depends on 4) The method of claim 4, wherein the stacked dilated 1D TCN blocks are stacked over each other such that a dilation factor goes on exponentially increasing as a function of the number of stacked layers. ### Claim 6 (depends on 5) The method of claim 5, wherein each 1D TCN block is a residual temporal gated linear unit (RT-GLU) incorporating TCN and residual learning into a new residual block. ### Claim 7 (depends on 1) The method of claim 1, wherein the decoder network transforms the weighted latent representation back into time-domain representations of the weighted latent representation using 1D transposed convolutions. ### Claim 8 (depends on 1) The method of claim 1, wherein the post-disaggregation adjustment method is used to refine the disaggregated native load and PV generation values around their predicted values by minimizing a total weighted squared error to ensure that a difference between the disaggregated native load and PV generation equals the net-load measurement. ### Claim 9 (depends on 1) The method of claim 1, wherein a discriminator is added to the neural networks regrouped as a generator to perform adversarial learning of energy disaggregation, wherein the input of the discriminator is shuffled between net-load measurements and net-loads determined based on an output of the generator, wherein the generator is composed by the encoder network, the feature extractor, the separator network, and the decoder network, wherein the input of the generator is the actual measurement data from CPOW units. ### Claim 10 (depends on 9) The method of claim 9, wherein the discriminator is a deeper network including convolutional layers for feature extraction, gated recurrent units (GRUs) for modeling temporal properties of net-loads, and sequentially stacked linear layers with decreasing dimensions for outputting a classification probability from the output of GRUs, wherein the probability indicates whether the input data is the actual measurement data from CPOW units or the generated data from the generator. ### Claim 11 (depends on 9) The method of claim 9, wherein the input of the discriminator is the feature space, and the discriminator is trained by two different sources of net-load data, one data source is a source domain with native load and PV generation labels, the other data source is a target domain without native load and PV generation labels, wherein the deviation between the source domain feature domain and the target domain feature space is minimized during the training. ### Claim 12 (depends on 11) The method of claim 11, wherein the discriminator consists of sequential fully connected layers with decreasing dimensions, wherein the final layer of the discriminator is a sigmoid activation function, wherein the output of the sigmoid function is a probability of the input data being from the source domain or target domain dataset. ### Claim 13 A energy disaggregation system for performing energy disaggregation of a distribution system-level net-load measurements of a distribution system using continuous-point-on-wave (CPOW) measurement units, comprising: a network interface controller configured to revieve measurement data of electric signals indicating voltage and current measurements of substations or feeders arranged in the distribution system; a memory configured to store instructions implementing an energy disaggregation method using neural networks including an encoder network, a feature extractor, a separator network, and a decoder network; and a processor coupled with the memory, wherein the processor is configured to execute the instructions that carry out steps of the energy disaggregation method, comprising: generating, using the encoder network and the measurement data of the electric signals, a compressed latent space representation from the net-load time series; converting the net-load time series into the time-frequency domain using a short-time Fourier transform (STFT) method; passing time-domain cotextual information with the converted time-frequency domain representation of net-load time series to the feature extractor trained to accurately extract harmonic features; estimating, using the separator network trained to assign weight to the compressed latent space representation, two weight matrices to be multiplied with an output from the encoder network and learning temporal features of a native load and a photovoltaic (PV) generation; transforming, using the decoder network, the weighted compressed latent representation corresponding the native load and the PV generation into time-domain representations from outputs of the encoder network and the separator network; and predicting, using a post-disaggregation adjustment method, the native load and the PV generation at distribution system-level from the transformed time-domain representations corresponding to the native load and PV generations. ### Claim 14 (depends on 13) The energy disaggregation system of claim 13, wherein the encoder network performs a one-dimensional (1D) convolution operation to transform the net-load time series into the compressed latent space representation of net-loads. ### Claim 15 (depends on 13) The energy disaggregation system of claim 13, wherein the feature extractor network identifies, using the STFT method, the harmonic features introduced in the distribution system by inverter-based PV generation. ### Claim 16 (depends on 13) The energy disaggregation system of claim 13, wherein the separator network uses a temporal convolutional network (TCN) to learn the learning, wherein the TCN includes sequential layers of stacked dilated 1D TCN blocks, wherein each of the stacked dilated 1D TCN blocks consists of a gated residual network. ### Claim 17 (depends on 16) The energy disaggregation system of claim 16, wherein the stacked dilated 1D TCN blocks are stacked over each other such that a dilation factor goes on exponentially increasing as a function of the number of stacked layers. ### Claim 18 (depends on 17) The energy disaggregation system of claim 17, wherein each 1D TCN block is a residual temporal gated linear unit (RT-GLU) incorporating TCN and residual learning into a new residual block. ### Claim 19 (depends on 13) The energy disaggregation system of claim 13, wherein the decoder network transforms the weighted latent representation back into time-domain representations of the weighted latent representation using 1D transposed convolutions. ### Claim 20 (depends on 13) The energy disaggregation system of claim 13, wherein the post-disaggregation adjustment method is used to refine the disaggregated native load and PV generation values around their predicted values by minimizing a total weighted squared error to ensure that the difference between the disaggregated native load and PV generation equals the net load measurement. ### Claim 21 (depends on 13) The energy disaggregation system of claim 13, wherein a discriminator is added to the neural network regrouped as a generator to perform adversarial learning of energy disaggregation, wherein the input of the discriminator is shuffled between net-load measurements and net-loads determined based on an output of the generator, wherein the generator is composed by the encoder network, the feature extractor, the separator network, and the decoder network, wherein the input of the generator is the actual measurement data from CPOW units. ### Claim 22 (depends on 21) The energy disaggregation system of claim 21, wherein the discriminator is a deeper network including convolutional layers for feature extraction, gated recurrent units (GRUs) for modeling temporal properties of net-loads, and sequentially stacked linear layers with decreasing dimensions for outputting a classification probability from the output of GRUs, wherein the probability indicates whether the input data is the actual measurement data from CPOW units or the generated data from the generator. ### Claim 23 (depends on 21) The energy disaggregation system of claim 21, wherein the input of the discriminator is the feature space, and the discriminator is trained by two different sources of net-load data, one data source is a source domain with native load and PV generation labels, the other data source is a target domain without native load and PV generation labels, wherein the deviation between the source domain feature domain and the target domain feature space is minimized during the training. ### Claim 24 (depends on 23) The energy disaggregation system of claim 23, wherein the discriminator consists of sequential fully connected layers with decreasing dimensions, wherein the final layer of the discriminator is a sigmoid activation function, wherein the output of the sigmoid function is a probability of the input data being from the source domain or target domain dataset.