Load decomposition method based on fusion feature data enhancement

Abstract

The invention discloses a load decomposition method based on fusion feature data enhancement, and the method comprises the steps: synchronously collecting the low-frequency power data of a bus end of an electrical loop of a building and the low-frequency power data of all electric equipment ends, and generating a confusion power sequence of similar equipment through Beta distribution mixing, so as to enhance the recognition capability of a model for power overlapping features; based on the power time sequence data, extracting a mutation feature, an equipment state feature and a time coding feature to construct a multi-dimensional feature vector; a CNN-BiLSTM double-branch neural network is adopted, spatial-temporal characteristics are fused through a dynamic weight attention mechanism, and equipment state classification and power decomposition tasks are jointly optimized. In practical application, bus end power data is input, and the operation state and power distribution of each electric device are obtained. According to the method, an adversarial training strategy and a gating feature fusion mechanism are innovatively introduced, the load decomposition performance in a complex power utilization scene is remarkably improved, and the method is particularly suitable for identification and power prediction of equipment with similar rated power.

Description

Technical Field

[0001] This invention relates to the field of power load monitoring technology, and in particular to a load decomposition method based on enhanced fusion feature data. Background Technology

[0002] With the development of smart grids and building energy-saving technologies, non-intrusive load monitoring technology has received widespread attention due to its advantages such as easy installation and low cost.

[0003] For example, Chinese patent document CN111428755A discloses a non-invasive load monitoring method that collects high-frequency current and voltage data, mines equipment features, calculates multi-dimensional feature values ​​that can characterize the operating imprint of electrical equipment, and finally establishes a random forest classifier model for classification and identification.

[0004] Chinese patent document CN106096726A discloses a non-intrusive load monitoring method, comprising: obtaining the sum of current and the sum of active power of each sampling point of the electrical load to be monitored within a set time period; using the sum of current and the sum of active power as input layer nodes of a neural network model; using the operating state of the electrical load as the output layer node of the neural network model; training the neural network model according to the sampling data obtained from monitoring within a set time period; and determining the operating state of each electrical load within the next set time period through the trained neural network model.

[0005] However, existing technologies mainly have the following problems:

[0006] 1. The identification accuracy is insufficient for devices with similar rated power. When multiple similar devices are running at the same time, traditional methods are difficult to accurately distinguish the power contribution of each device.

[0007] 2. In complex power consumption scenarios, equipment start-up and shutdown and power fluctuations lead to overlapping load characteristics, and existing models have limited ability to identify power overlap characteristics;

[0008] 3. Most methods focus only on a single task such as power decomposition or state recognition, lacking a multi-feature joint learning mechanism, resulting in poor overall performance;

[0009] 4. High data collection requirements: A large amount of labeled data is needed during the training phase, which is difficult to meet in practical applications. Summary of the Invention

[0010] This invention provides a load decomposition method based on fusion feature data augmentation, which significantly improves the accuracy and robustness of load decomposition through innovative data augmentation strategies and deep learning architecture.

[0011] A load decomposition method based on fused feature data augmentation includes the following steps:

[0012] (1) Collect the total power sequence of the low frequency at the bus terminal of the electrical circuit of the target building, and simultaneously collect the power sequence of the low frequency at the terminal of each electrical device;

[0013] (2) Select similar devices with a rated power close to that of the target device, generate a mixed power sequence through Beta distribution, and randomly add the mixed power sequence to the total power sequence at the bus end to obtain the total power mixed timing data;

[0014] (3) Extract mutation features, equipment status features and time coding features based on total power mixed time series data to form a multi-dimensional feature vector, and divide the training set, validation set and test set according to a certain ratio;

[0015] (4) Construct a CNN-BiLSTM dual-branch neural network, fuse spatiotemporal features through a dynamic weight attention mechanism, construct a multi-task loss function, and train the model parameters using an adversarial training strategy;

[0016] (5) During the application process, input the power data at the bus end and output the power prediction value and operating status classification of the target device.

[0017] This invention solves the problems of difficulty in identifying similar devices, low distinguishability of power overlap features, and insufficient multi-task learning in the prior art.

[0018] The specific process of step (1) is as follows:

[0019] (1-1) Using current and voltage sensors with a sampling frequency of ≥4kHz, high-frequency voltage and current waveforms of the target building's electrical circuit bus terminal and each electrical equipment terminal are collected synchronously.

[0020] (1-2) After acquiring the high-frequency voltage and current waveforms, the terminal equipment calculates the active power using the discrete integral formula:

[0021]

[0022] Where N is the number of sampling points per cycle, and u and i represent the voltage and current at a certain moment, respectively; then the average downsampling method is used to downsample the data to 1 / 5Hz-5Hz;

[0023] (1-3) The terminal device uploads the data packet to the server every 60 seconds. The data packet contains the time-aligned power sequence of the bus end and the device end. The server stores the data as structured data according to the building ID and timestamp.

[0024] In step (1-1), when collecting the high-frequency voltage and current waveforms of each electrical device, at least one high-frequency voltage and current waveform of the target device should be included. Each collection point should use a unified clock source to ensure that the time alignment error is ≤100ms. The target device and other devices should be kept running simultaneously to provide complete load characteristic data.

[0025] The specific process of step (2) is as follows:

[0026] (2-1) Select equipment whose average operating power in each operating cycle is between 0.8 and 1.2 times that of the target equipment as similar equipment;

[0027] (2-2) For the target device and similar devices, randomly select their operating segments and generate a mixed power sequence P according to λ~Beta(0.4,0.4). mix =λP1+(1-λ)P2, resulting in a more similar and confusing data segment;

[0028] (2-3) Randomly add the obtained confused data fragments to the total power sequence at the bus end to obtain the total power mixed timing data.

[0029] The specific process of step (3) is as follows:

[0030] (3-1) Calculate the first difference ΔP of the total power mixed time series data. t =P t -P t-1 , indicating the characteristics of mutation data;

[0031] (3-2) For the power sequence data of the target equipment, the k-means++ algorithm is used to divide the operating state, and the k value is determined by the silhouette coefficient optimization; the power sequence of the target equipment is divided into a multi-state sequence as the equipment state feature;

[0032] (3-3) Construct time-coding features, including sinusoidal function codes for intraday, intraweek, and intrayear cycles, as well as markers for peak hours, holidays, and weekdays:

[0033] T t =PE hour ⊕PE week ⊕PE year ⊕[I peak ,I holiday ,I workday ]

[0034] In the formula, Indicates intraday cycle encoding, Indicates the intra-week cycle code. Indicates the intra-year cycle code. Indicates peak time period markers. Indicates holiday markers, Indicates a workday marker, where It is an indicator function (it takes the value 1 if the condition is true, and 0 otherwise), ∧ represents logical AND, and ∨ represents logical OR. The symbol ⊕ represents logical NOT, and ⊕ represents vector concatenation. The final encoded dimension is d. t =3×2+3=9 dimensions;

[0035] (3-4) Based on a 6:2:2 ratio, the data is divided into training set, validation set and test set; X of each dataset represents the characteristics of the bus end, namely total power value, mutation data characteristics and time coding characteristics; Y of each dataset represents the characteristics of the target device end, namely the target device power value and device status characteristics.

[0036] The specific process of step (4) is as follows:

[0037] (4-1) Divide the input data X into sliding window samples according to the window length L, and use one sample at a time;

[0038] (4-2) Perform multi-feature joint encoding on the input total power mixed time series data to obtain fusion features with time-series dependencies;

[0039] (4-3) The device status is identified by the convolutional neural network branch of the CNN-BiLSTM dual-branch neural network, and the device status at each time step is output.

[0040] (4-4) The power decomposition of the fused features is performed by the long short-term memory network branch of the CNN-BiLSTM dual-branch neural network, and the device power allocation at each time step is output.

[0041] (4-5) An adversarial training strategy is adopted to jointly optimize the state recognition and power decomposition results, calculate the multi-task loss function and perform backpropagation.

[0042] The specific process of step (4-2) is as follows:

[0043] (4-2-1) Construct multi-feature input data, including a sequence of total power values. Power mutation value sequence and timestamp encoding

[0044] (4-2-2) Construct the power-time joint feature space E respectively pt and mutation-time joint feature space E Δt :

[0045]

[0046] In the formula, d model The model dimension is represented by LayerNorm, which represents the layer normalization operation, Linear, which represents the linear transformation, and Concat, which represents the concatenation operation.

[0047] (4-2-3) The two joint feature spaces are fused using a gating mechanism to obtain the fused feature Z:

[0048]

[0049] In the formula, g represents the gating signal, σ represents the Sigmoid activation, which maps the weight scores to the interval [0,1], and ⊙ represents element-wise multiplication.

[0050] The specific process of step (4-3) is as follows:

[0051] (4-3-1) Input the fused feature Z into the residual convolutional network and calculate the feature F. state :

[0052] F state =ResBlock(Conv1D(Z,filters=d model ,kernel=3))

[0053] In the formula, ResBlock represents the residual block, Conv1D represents the one-dimensional convolution, filters represents the hidden layer size, and kernel represents the convolution kernel width;

[0054] (4-3-2) The target device state prediction probability is output by the classifier.

[0055]

[0056] In the formula, Softmax represents the normalized exponential function;

[0057] Where N represents the number of states of the target device, excluding the off state 0.

[0058] The specific process of step (4-4) is as follows:

[0059] (4-4-1) Input the fused feature Z into the bidirectional LSTM network and calculate the feature H:

[0060]

[0061] In the formula, BiLSTM represents a bidirectional long short-term memory network, and units represents the hidden state dimension;

[0062] (4-4-2) Output the target equipment power prediction results through the time series regressor.

[0063]

[0064] In steps (4-5), the specific process of calculating the multi-task loss function is as follows:

[0065] (4-5-1) An improved focus loss function is used to handle multi-state classification, and the state classification loss is calculated. The formula is:

[0066]

[0067] In the formula, S t This represents the actual state of the target device at time t. This represents the probability of the target device being in state n at time t.

[0068] in, γ = 2 is a hyperparameter;

[0069] (4-5-2) Improve event detection accuracy by gradient matching at state transition moments, and calculate start / stop boundary loss. The formula is:

[0070]

[0071] in, ||·|| represents the L2 norm (Euclidean distance), which calculates the distance between two vectors;

[0072] (4-5-3) Calculate the power regression loss using the weighted MSE obtained from device startup sensing. The formula is:

[0073]

[0074] In the formula, Y t This represents the actual power of the target device at time t. This represents the predicted power of the target device at time t;

[0075] (4-5-4) Calculate the regression accuracy penalty term at the startup time of the computing device. The formula is:

[0076]

[0077] in, This represents the set of all startup times for device n. The difference operator representing adjacent time steps;

[0078] (4-5-5) Finally, calculate the multi-task loss function:

[0079]

[0080] Where λ1, λ2, λ3, and λ4 are balance coefficients.

[0081] Compared with the prior art, the present invention has the following beneficial effects:

[0082] 1. This invention generates confused power sequences of similar devices by mixing Beta distributions, which significantly enhances the model's ability to identify power overlap features, and is particularly suitable for distinguishing devices with similar rated power;

[0083] 2. This invention innovatively constructs a multi-dimensional feature vector, including mutation features, equipment status features, and time-encoded features, to comprehensively capture load feature information;

[0084] 3. This invention adopts a CNN-BiLSTM dual-branch neural network architecture, and achieves effective fusion of spatiotemporal features through a dynamic weight attention mechanism, while optimizing the tasks of device state classification and power decomposition.

[0085] 4. This invention introduces an adversarial training strategy and a gating feature fusion mechanism to improve the robustness of the model in complex power consumption scenarios;

[0086] 5. In practical applications, this invention only requires inputting bus power data to accurately predict the operating status and power allocation of each device, greatly reducing data acquisition requirements. Attached Figure Description

[0087] Figure 1 This is a flowchart of a load decomposition method based on fused feature data enhancement according to an embodiment of the present invention.

[0088] Figure 2 This is a schematic diagram of the data augmentation process in an embodiment of the present invention.

[0089] Figure 3 This is a schematic diagram of the structure of the CNN-BiLSTM dual-branch neural network in an embodiment of the present invention. Detailed Implementation

[0090] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0091] like Figure 1 As shown, a load decomposition method based on fused feature data augmentation includes the following steps:

[0092] S10. Collect and process low-frequency data: Collect the total low-frequency power value of the target building's electrical circuit bus terminal and synchronously collect the low-frequency power value of each electrical device terminal.

[0093] S20. Data augmentation for similar devices: Select devices with similar rated power to the target device, and generate confused power sequences through Beta distribution mixing to force the model to learn power overlap features;

[0094] S30. Feature Fusion: Extract abrupt change features, equipment status features, and time coding features based on power time series data to form a multi-dimensional feature vector, and divide it into training set, validation set, and test set according to a certain ratio;

[0095] S40. Multi-feature joint learning: Construct a CNN-BiLSTM dual-branch neural network, fuse spatiotemporal features through a dynamic weight attention mechanism, use multi-feature data as input to train model parameters, and output the power prediction value and operating status classification of the target device.

[0096] Furthermore, the specific implementation of step S10 is as follows:

[0097] S101. Using current and voltage sensors with a sampling frequency of ≥4kHz, synchronously acquire high-frequency voltage and current waveforms between the target building's electrical circuit bus terminal and each electrical equipment terminal.

[0098] S102. After acquiring the high-frequency voltage and current waveforms, the terminal equipment calculates the active power using the discrete integral formula:

[0099]

[0100] Where N is the number of sampling points per cycle, and u and i represent the voltage and current at a certain moment, respectively. The data is then downsampled to 1 / 5Hz-5Hz using an averaging downsampling method.

[0101] S103. Terminal devices upload data packets to the server every 60 seconds. The data packets contain time-aligned power sequences for the bus and device ends. The server stores the data as structured data by building ID and timestamp.

[0102] Furthermore, the data acquisition in step S101 must meet the following requirements: during the training phase, high-frequency voltage and current waveforms of the bus end and at least one target device and other electrical equipment ends are acquired synchronously. Each acquisition point uses a unified clock source to ensure that the time alignment error is ≤100ms, and the target device and other devices must maintain a simultaneous operating state to provide complete load characteristic data.

[0103] Figure 2 This is a schematic diagram of the data augmentation process in this invention. (For example...) Figure 2 As shown, the specific implementation method of step S20 is as follows:

[0104] S201. Select equipment whose average operating power is between 0.8 and 1.2 times that of the target equipment in each operating cycle as similar equipment;

[0105] S202. For the target device and devices with similar power, randomly select their operating segments and generate a mixed power sequence P according to λ~Beta(0.4,0.4). mix =λP1+(1-λ)P2, by obtaining confused data fragments with higher similarity, the model is forced to learn fuzzy boundary features;

[0106] S203. Obtain the obfuscated data fragment and randomly add it to the power value data at the bus end;

[0107] S204. During training, 15% of the batches are randomly selected for adversarial training. For each selected batch, 50% of the samples are replaced with adversarial samples. The FGSM method is used to generate perturbations to improve the model's robustness to noise.

[0108]

[0109] Where ∈ represents the perturbation coefficient. Let θ represent the gradient of the loss function J with respect to the input power sequence x, and let θ represent the weight parameters of the load decomposition model.

[0110] Furthermore, in step S201, if no device meets the conditions, the operating power data of the candidate device is linearly scaled and adjusted to match its power characteristics with those of the target device.

[0111] Furthermore, the specific implementation of step S30 is as follows:

[0112] S301. Calculate the first-order difference value ΔP of the power sequence data at the bus terminal. t =P t -P t-1 This indicates abrupt changes in data characteristics. For the bus side, abrupt changes in data characteristics, compared to power sequence data, represent stable data characteristics, avoiding numerical fluctuations caused by variations in the number of operating devices.

[0113] S302. For the power sequence data of the target device, the k-means++ algorithm is used to divide the operating state. The k value is determined by the silhouette coefficient optimization. The power sequence of the target device is divided into a multi-state sequence as an auxiliary feature of the device data.

[0114] S303. Construct a composite time feature vector, including sine function encodings of intraday (24h), intraweek (7d), and intrayear (366d) cycles, as well as peak hours, holiday and weekday markers:

[0115] T t =PE hour ⊕PE week ⊕PE year⊕[I peak ,I holiday ,I workday ]

[0116] Enhance the model's ability to perceive the cyclical patterns of electricity consumption behavior;

[0117] S304. Divide the data into training, validation, and test sets according to a 6:2:2 ratio. X represents the bus-side characteristics of each dataset, including total power value, abrupt data characteristics, and time-coded characteristics. Y represents the device-side characteristics of each dataset, including target device power value and target device status characteristics.

[0118] Furthermore, step S40 is specifically implemented as follows:

[0119] S401. Divide the input data X into sliding window samples according to the window length L, and use one sample at a time.

[0120] S402. Perform multi-feature joint encoding on the input mixed power time series data to obtain a feature representation with time-dependent relationships;

[0121] S403. Use convolutional neural network branches to identify the device status of the encoded features and output the device status at each time step.

[0122] S404. Power decomposition of encoded features is performed through branches of the Long Short-Term Memory network to output the device power allocation at each time step.

[0123] S405. Combine the results of state identification and power decomposition, calculate the multi-task loss function, and perform backpropagation.

[0124] Figure 3 This is a schematic diagram of the CNN-BiLSTM dual-branch neural network in this invention.

[0125] Furthermore, the specific implementation of step S402 is as follows:

[0126] S4021. Construct multi-feature input data, including a total power value sequence. Power mutation value sequence and timestamp encoding

[0127] S4022. Construct the power-time joint feature space and the mutation-time joint feature space respectively:

[0128]

[0129] S4023. Merge the two joint feature spaces using a gating mechanism:

[0130]

[0131] Furthermore, the specific implementation of step S403 is as follows:

[0132] S4031. Input the fused feature Z into the residual convolutional network:

[0133] F state =ResBlock(Conv1D(Z,filters=d model ,kernel=3))

[0134] S4032. Output device status via classifier:

[0135]

[0136] Where N represents the number of states of the target device, excluding the off state 0.

[0137] Furthermore, the specific implementation of step S404 is as follows:

[0138] S4041. Input the fused feature Z into the bidirectional LSTM network:

[0139]

[0140] S4042. Output the prediction results through the time series regressor:

[0141]

[0142] Furthermore, the specific implementation of S405 is as follows:

[0143] S4051. Use an improved focus loss function to handle multi-state classification and calculate the state classification loss:

[0144]

[0145] in, γ = 2 is a hyperparameter;

[0146] S4052. Improve event detection accuracy by gradient matching at state transition moments, and calculate start / stop boundary losses.

[0147]

[0148] in,

[0149] S4053. Calculate power regression loss using weighted MSE obtained from device startup sensing.

[0150]

[0151] S4054, Regression accuracy penalty term at the startup time of computing devices

[0152]

[0153] in, This represents the set of all startup times for device n. The difference operator representing adjacent time steps;

[0154] S4055. Finally, calculate the total loss function:

[0155]

[0156] Where λ1, λ2, λ3, and λ4 are balance coefficients.

[0157] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A load decomposition method based on fusion feature data enhancement, characterized in that, Comprise the following steps: (1) Collect the total power sequence of the target building electrical circuit bus end low frequency, and synchronously collect the power sequence of each power equipment end low frequency; (2) Select similar equipment with a power close to the rated power of the target equipment, generate mixed power sequence by Beta distribution mixing, randomly add the mixed power sequence to the total power sequence of the bus end, and obtain total power mixed time series data; (3) Extract mutation features, equipment state features and time coding features based on the total power mixed time series data, form a multi-dimensional feature vector, and divide the training set, validation set and test set according to a certain proportion; (4) Construct a CNN-BiLSTM double-branch neural network, fuse the space-time features through a dynamic weight attention mechanism, construct a multi-task loss function, and train the model parameters using an adversarial training strategy; (5) In the application process, input the bus end power data, and output the power prediction value and running state classification of the target equipment.

2. The load decomposition method based on fusion feature data enhancement according to claim 1, characterized in that, The specific process of step (1) is as follows: (1-1) Use current and voltage sensors with a sampling frequency of ≥4kHz to synchronously collect high-frequency voltage and current waveforms of the target building electrical circuit bus end and each power equipment end; (1-2) After collecting the high-frequency voltage and current waveforms, the terminal device calculates the active power by using the discrete integral formula: Wherein, N is the number of sampling points per cycle, u and i represent the voltage and current at a certain time, respectively; Then use the average downsampling method to reduce the data to 1 / 5Hz-5Hz; (1-3) The terminal device uploads to the server every 60 seconds, and the data packet contains the time-aligned bus end and equipment end power sequence, and the server stores it as structured data according to the building ID and timestamp.

3. The load decomposition method based on fusion feature data enhancement according to claim 2, characterized in that, In step (1-1), when collecting the high-frequency voltage and current waveforms of each power equipment end, at least the high-frequency voltage and current waveforms of the target equipment are included, and a unified clock source is used to ensure that the time alignment error is ≤100ms, and the target equipment and other equipment need to maintain a simultaneous running state to provide complete load feature data. 4.The load decomposition method based on fusion feature data enhancement according to claim 1, characterized in that, The specific process of step (2) is as follows: (2-1) Select devices with a ratio of the average running power to the target device within each running period in the range of 0.8 to 1.2 times as similar devices; (2-2) For the target device and the similar device, randomly select two running fragments, and generate a mixed power sequence P according to λ ~ Beta(0.4, 0.4) mix = λP1+(1-λ)P2, obtain a more similar confused data fragment; wherein, P1 represents the power of the target device, and P2 represents the power of the similar device; (2-3) Randomly add the obtained mixed data segment to the total power sequence of the bus end to obtain the total power mixed time series data.

5. The load decomposition method based on fusion feature data enhancement according to claim 1, characterized in that, The specific process of step (3) is as follows: (3-1) Calculate the first difference value ΔP of the total power mixed time series data t = P t - P t-1 , indicates the mutation data characteristics; (3-2) Divide the running state of the target equipment power sequence data using the k-means++ algorithm, and the value of k is determined by optimizing the silhouette coefficient; Divide the power sequence of the target equipment into state sequences of multiple states as equipment state features; (3-3) Construct time coding features, including sine function coding of intra-day cycle, intra-week cycle and intra-year cycle, and high peak period, holiday and weekday markers: T t = PE hour ⊕ PE week ⊕ PE year ⊕ [I peak , I holiday , I workday ] where PE hour denotes the intra-day periodicity encoding, PE week denotes the intra-week periodicity encoding, PE year denotes the intra-year periodicity encoding, I peak denotes the peak hour marker, I holiday denotes the holiday marker, I workday denotes the weekday marker, and denotes the vector concatenation operation. (3-4) According to the 6:2:2 ratio, divide the data into training set, validation set and test set; The X of each data set represents the features of the bus end, i.e. total power value, mutation data feature and time coding feature; The Y of each data set represents the equipment end features of the target equipment, i.e. target equipment power value and equipment state feature.

6. The load decomposition method based on fusion feature data enhancement according to claim 5, characterized in that, The specific process of step (4) is: (4-1) Divide the input data X into sliding window samples according to the window length L, and use one sample at a time; (4-2) Jointly encode the input total power mixed time series data to obtain fused features with time dependence; (4-3) Perform device state recognition on the fused features through the convolutional neural network branch of the CNN-BiLSTM dual-branch neural network, and output the device state at each time; (4-4) Perform power decomposition on the fused features through the long short-term memory network branch of the CNN-BiLSTM dual-branch neural network, and output the device power allocation at each time; (4-5) Jointly optimize the state recognition and power decomposition results using an adversarial training strategy, calculate the multi-task loss function, and perform backpropagation.

7. The load decomposition method based on fusion feature data enhancement according to claim 6, characterized in that, The specific process of step (4-2) is: (4-2-1) Constructing multi-feature input data, including total power value sequence Power mutation value sequence And timestamp encoding (4-2-2) Construct power-time joint feature space E pt and mutation-time joint feature space E Δt : In the formula, d model denotes the model dimension, LayerNorm denotes a layer normalization operation, Linear denotes a linear transformation, and Concat denotes a concatenation operation. (4-2-3) Fuse the two joint feature spaces through a gating mechanism to obtain fused features Z: where g represents the gating signal, σ represents the Sigmoid activation that maps the weight score to the [0, 1] interval, and represents element-wise multiplication. 8.The load decomposition method based on fusion feature data enhancement according to claim 6, characterized in that, The specific process of step (4-3) is: (4-3-1) Input the fusion feature Z into the residual convolutional network, calculate the feature F state : F state = ResBlock(Conv1D(Z, filters=d model , kernel=3)) where ResBlock represents the residual block, Conv1D represents one-dimensional convolution, filters represent the size of the hidden layer, and kernel represents the width of the convolution kernel. (4-3-2) Outputting a target device state prediction probability by a classifier where Softmax represents the normalized exponential function. where N represents the number of states of the target device, excluding the off state 0. 9.The load decomposition method based on fusion feature data enhancement according to claim 6, characterized in that, The specific process of step (4-4) is: (4-4-1) Input the fused features Z into the bidirectional LSTM network to calculate the features H: where BiLSTM represents the bidirectional long short-term memory network, and units represent the hidden state dimension. (4-4-2) Outputting target device power prediction results by time series regressor 10. The load decomposition method based on fusion feature data enhancement according to claim 6, characterized in that, In step (4-5), the specific process of calculating the multi-task loss function is: (4-5-1) The improved focal loss function is used to process multi-state classification, and the state classification loss is calculated The formula is: In the formula, S t denotes the real state of the target device at time t, denotes the probability of the state n of the target device at time t; wherein, y = 2 is a hyperparameter; (4-5-2) Improve the event detection accuracy by matching the gradient of the state switching moment, calculate the start-stop boundary loss The formula is: wherein ||•|| denotes the L2 norm, which calculates the distance between two vectors; (4-5-3) Weighted MSE with device start awareness, compute power regression loss The formula is: In the formula, Y t represents the real power of the target device at time t, represents the predicted power of the target device at time t; (4-5-4) Regression accuracy penalty term for the start time of the computing device The formula is: wherein, denotes the set of all start times of device n, denotes the difference operator for adjacent time steps; (4-5-5) Finally, calculate the multi-task loss function: where λ1, λ2, λ3, and λ4 are balance coefficients.

Images