Industrial load decomposition method based on harmonic coupling and nonlinear feature fusion

Abstract

The invention provides an industrial load decomposition method based on harmonic coupling and nonlinear feature fusion, and belongs to the technical field of intelligent power grid monitoring. The method comprises the following steps: acquiring an electrical signal at a main power supply of an industrial load; basic electrical characteristics are extracted based on the electrical signals, and a basic characteristic set is constructed; geometric mean features are determined through the basic electrical features, and a geometric mean feature set and a multi-time scale geometric mean feature set are constructed; combining the basic feature set, the multi-time scale geometric mean feature set and the geometric mean feature set to construct a fusion feature set; training a long-short-term memory network by using the fused feature set; and performing non-intrusive industrial load energy consumption monitoring by using the trained long-short-term memory network. According to the method, the long-short-term memory network is trained by fusing the feature set, and the trained long-short-term memory network set is utilized to carry out non-intrusive load monitoring, so that the energy consumption of each load in the industrial load can be more accurately detected, decomposed and evaluated.

Description

Technical Field

[0001] This invention relates to the field of smart grid monitoring technology, and in particular to an industrial load decomposition method based on the fusion of harmonic coupling and nonlinear characteristics. Background Technology

[0002] As energy demand from industrial loads increases, optimal energy usage strategies are needed for effective energy management in both residential and industrial loads. Traditional methods involve real-time monitoring of energy consumption across the entire industrial load and sending data to consumers at user-defined intervals or continuously. This conventional approach to providing real-time energy consumption feedback to consumers can save approximately 9% of total energy consumption. However, the lack of detail regarding the electrical infrastructure used in industrial loads means these methods have limited impact on consumer energy usage behavior. Therefore, load / equipment-level energy monitoring is needed to understand consumer energy consumption patterns. However, this approach requires sensors for each load, which is costly and impractical. Therefore, non-intrusive load monitoring (NILM) has been introduced to monitor the electricity utility by measuring the total energy at the industrial load inlet.

[0003] Non-intrusive load monitoring (NILM) aims to detect loads and decompose the energy consumption of individual loads by extracting features of aggregated energy at a given time to capture the characteristic electrical signals of loads during transient and steady-state operation. The extracted features are then fed into machine learning and deep learning algorithms for load classification and energy decomposition. However, traditional NILM systems typically rely on basic electrical characteristics (such as voltage, current, and power), which cannot adequately capture the unique behavior of loads. When dealing with multiple similar or identical loads, it is difficult to distinguish their energy consumption characteristics, leading to inaccurate detection. Existing NILM methods rely on basic features such as active / reactive power, making it difficult to distinguish industrial equipment with similar harmonic characteristics (such as multiple frequency converters), and a single time scale cannot accommodate transient responses (such as motor starting) and long-cycle energy consumption patterns (such as production shifts). Furthermore, traditional geometric mean calculation involves multiplying high-dimensional features, resulting in poor numerical stability and low computational efficiency.

[0004] Therefore, there is a need for an industrial load decomposition method based on the fusion of harmonic coupling and nonlinear characteristics. Summary of the Invention

[0005] In view of this, the present invention provides an industrial load decomposition method based on harmonic coupling and nonlinear feature fusion, which constructs a geometric feature set to train a long short-term memory network, and uses the trained long short-term memory network to input the signal collected at the main electricity meter to achieve non-intrusive load monitoring.

[0006] Therefore, the present invention provides the following technical solution: An industrial load decomposition method based on the fusion of harmonic coupling and nonlinear characteristics includes: Collect electrical signals at the main power source of the industrial load; Based on the electrical signals, basic electrical features are extracted, and a basic feature set is constructed; Geometric mean features are determined by basic electrical characteristics, and a geometric mean feature set and a multi-time-scale geometric mean feature set are constructed. A fused feature set is constructed by combining the basic feature set, the multi-time-scale geometric mean feature set, and the geometric mean feature set. Train a long short-term memory network using a fused feature set; Non-invasive industrial load energy consumption monitoring is performed using a trained long short-term memory network.

[0007] Furthermore, the electrical signal includes: The electrical signals of the R-phase and Y-phase of the three-phase current signal.

[0008] Furthermore, the basic feature set includes: Active power, reactive power, apparent power, power factor, harmonic content.

[0009] Further, the geometric mean feature set includes: The geometric mean of active power and reactive power, the geometric mean of reactive power and apparent power, the geometric mean of active power and current, the geometric mean of apparent power and current, the geometric mean of active power and harmonic content, the geometric mean of reactive power and harmonic content, the geometric mean of apparent power and harmonic content, the geometric mean of active power, reactive power and apparent power, and the geometric mean of active power, reactive power, apparent power and harmonic content.

[0010] Furthermore, it also includes preprocessing the electrical signals: Remove outliers and noise, and fill in or delete missing data points.

[0011] Furthermore, the multi-timescale geometric mean feature set includes: Transient-scale geometric mean, steady-state-scale geometric mean, long-period-scale geometric mean, and cross-scale time geometric mean.

[0012] Furthermore, the method of training the long short-term memory network using a fused feature set includes: The Long Short-Term Memory (LSTM) network takes a fused feature set as input and outputs the energy consumption of each load.

[0013] Further, the geometric mean characteristic is calculated through logarithmic transformation, including: The formula for calculating the geometric mean is:

[0014] in, This represents the basic electrical characteristics; n represents the total number of basic electrical characteristics involved in the calculation. Calculate the logarithmic mean:

[0015]

[0016] in, It represents the geometric mean.

[0017] Advantages and positive effects of the present invention: This method trains a long short-term memory network by fusing feature sets and uses the trained long short-term memory network set for non-invasive load monitoring. This can more accurately detect and decompose the energy consumption of each load in industrial loads, solving the problem that existing NILM technology is insufficient in decomposing industrial loads under similar load conditions.

[0018] The feature set fused in this method includes: Introducing harmonic content as a load fingerprint feature when analyzing basic electrical characteristics provides a more comprehensive analysis. Harmonics can serve as additional clues to distinguish different load characteristics, especially when load power signals are similar. Analysis of harmonic components helps identify inefficient equipment and operations, enabling measures to improve energy efficiency. Furthermore, harmonic analysis can reveal potential equipment problems and assess grid issues leading to system instability or failure, such as overheating, aging, or damage. This facilitates preventative maintenance and fault diagnosis, thus contributing to the development of more effective energy management and energy-saving strategies.

[0019] Simultaneously, by utilizing a multi-timescale geometric average feature set, a multi-timescale geometric average feature is proposed to simultaneously capture millisecond-level transients and hourly-level trends. This resolves the contradictory requirement in industrial load monitoring to capture both millisecond-level transient details (such as harmonic abrupt changes during switching) and hourly-level operating trends. It achieves effective extraction and collaborative fusion of multi-timescale features from transient to long-cycle perspectives.

[0020] The logarithmic transformation method is used when calculating the geometric mean, which improves the stability of numerical calculation, especially when dealing with very large or very small values. Most importantly, in a parallel computing environment, addition and exponentiation operations are usually faster than multiplication operations, thus improving computational efficiency and shortening computer processing time. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the non-intrusive industrial load energy consumption monitoring method based on geometric feature fusion according to Embodiment 1 of the present invention; Figure 2 This is a logic diagram of the non-intrusive industrial load energy consumption monitoring method based on geometric feature fusion in Embodiment 2 of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] This invention provides an industrial load decomposition method based on harmonic coupling and nonlinear feature fusion. By training a long short-term memory network through a fusion feature set, and using the trained long short-term memory network set for non-intrusive load monitoring, the energy consumption of each load in the industrial load can be detected and decomposed more accurately.

[0026] Example 1 Combination Figure 1 As shown, an industrial load decomposition method based on the fusion of harmonic coupling and nonlinear characteristics includes: Step 1: Collect electrical signals at the main power source of the industrial load and extract electrical features as a basic feature set.

[0027] Step 2: Calculate the geometric mean based on the basic electrical characteristics, construct the geometric mean feature set and the multi-time-scale geometric mean feature set, and obtain the fused feature set by combining the geometric mean feature set, the basic feature set and the multi-time-scale geometric mean feature set.

[0028] Step 3: Train a Long Short-Term Memory (LSTM) network using the fused feature set. Evaluate the performance of the LSTM model using an independent test set, and use root mean square error, mean absolute error, and energy decomposition accuracy to measure the accuracy of energy decomposition, and adjust the model parameters based on the test results.

[0029] Step 4: Utilize the trained LSTM model to monitor and decompose the real-time energy consumption of various similar loads in the industrial load, providing detailed energy consumption data for energy management and continuous monitoring of system operation. Collect feedback data to further train and optimize the LSTM model, ensuring the system maintains high performance in constantly changing industrial load environments.

[0030] Example 2 Combination Figure 2 As shown, an industrial load decomposition method based on the fusion of harmonic coupling and nonlinear characteristics includes: Step 1: Collect basic electrical characteristics at the main power source of the industrial load, including: active power, reactive power, apparent power, power factor, RMS current and harmonic content.

[0031] 1) At the main power source of the industrial load, collect the electrical signals of the R phase and Y phase of the three-phase current signal.

[0032] 2) Extracting basic electrical characteristics from electrical signals: Construct a mathematical model of instantaneous power under steady operating conditions:

[0033] Wherein, the instantaneous voltage of a certain electrical appliance at time t is The instantaneous current is , This refers to instantaneous power.

[0034] Determine the active power:

[0035] Determining the effective voltage value based on instantaneous power. and the effective value of current :

[0036]

[0037] Within one cycle, the apparent power S is obtained by multiplying the effective value of the voltage and the effective value of the current:

[0038] In alternating current circuits, the mathematical relationships based on the fundamental electrical characteristics are... Determine the power factor:

[0039] in, The power factor.

[0040] Calculate the harmonic content: The time-domain signal is converted into a frequency-domain signal using Fourier transform:

[0041] in, It is a time-domain signal. It is a frequency domain representation. It refers to frequency.

[0042] Harmonic component calculation involves extracting the amplitudes of the fundamental frequency (Frequency) and each harmonic (Odds) from the frequency spectrum.

[0043] in, It is the amplitude of the nth harmonic. It is the fundamental frequency.

[0044] Harmonic content is calculated as the ratio of each harmonic to the fundamental frequency, usually expressed as a percentage, and the formula is as follows:

[0045] THD stands for Total Harmonic Distortion. It is the fundamental frequency amplitude.

[0046] Using P (active power), Q (reactive power), S (apparent power), PF (power factor), and H (harmonic content) as input features, a basic feature set BF is constructed:

[0047] Step 2: Construct a fused feature set based on the geometric mean of the basic electrical characteristics.

[0048] Since each basic electrical characteristic contains some important features of the load, these important basic electrical characteristics can be used to develop energy classification systems. In the field of machine learning, the geometric mean, used to capture the interactions between multiple variables, can therefore be used to expand the feature set.

[0049] For regression analysis, the geometric mean of multiple variables can be obtained by calculating the square root of their product, using the following formula:

[0050] The geometric mean represents the nth root of the product of the input features. Essentially, it combines multiple features into a single metric to represent the overall relationship between them.

[0051] S21. The geometric mean characteristics in this embodiment include: By calculating the active power P and reactive power Q ( Reactive power Q and apparent power S ), active power P and current ( Apparent power S and current ( ), active power P and harmonic content H ( Reactive power Q and harmonic content H Apparent power S and harmonic content H The geometric mean of active power P, reactive power Q, and apparent power S. Active power (P), reactive power (Q), apparent power (S), and harmonic content (H) ).

[0052] Specifically: Active power and reactive power are combined into a comprehensive index, which reflects their overall relationship. The formula is as follows:

[0053] Reactive power and apparent power are combined into a comprehensive index, which reflects the overall relationship between them. The formula is expressed as:

[0054] The formula for combining active power and apparent power into a single comprehensive index, reflecting the overall relationship between them, is expressed as follows:

[0055] The overall relationship between active power P and current I is expressed by the following formula:

[0056] The overall relationship between apparent power S and current I is expressed by the formula:

[0057] The overall relationship between active power P and harmonic content H is expressed by the formula:

[0058] The overall relationship between reactive power Q and harmonic content H is expressed by the formula:

[0059] The overall relationship between apparent power S and harmonic content H is expressed by the formula:

[0060] The formula for calculating the square root of the product of active power, reactive power, and apparent power is as follows:

[0061] To capture more complex relationships between variables, specifically the overall relationship between active power, reactive power, apparent power, and harmonic content, the formula is expressed as:

[0062] S22. Because the current feature set only uses instantaneous values, it cannot capture the typical multi-timescale effects of industrial load operation, such as: transient processes (10ms level): motor starting current surges, inverter switching transients; steady-state processes (1s level): power fluctuations during continuous operation; and long-cycle changes (1h level): load mode switching caused by production shifts. Therefore, a multi-timescale geometric mean feature set is constructed to solve the contradictory needs of "capturing both millisecond-level transients and predicting hourly trends" in industrial scenarios.

[0063] Multi-timescale geometric mean feature set, including: Transient Scale Geometric Mean Steady-state scale geometric mean and long-period geometric mean .

[0064] The formula for calculating the transient scale geometric mean is:

[0065] in, Indicates instantaneous active power. Indicates transient harmonic content, Indicates the rate of change of frequency. The reference frequency is 50Hz, which is used in this embodiment.

[0066] Calculated using wavelet ridges The formula is:

[0067] In this embodiment, The calculation was performed using a sampling frequency set to 8kHz and a 10ms sliding window.

[0068] in, , This represents the instantaneous voltage value, specifically the instantaneous voltage value measured at the i-th sampling time. This represents the instantaneous current value, that is, the instantaneous current value sampled in strict synchronization.

[0069] In this embodiment, Extracting the 2nd to 15th harmonics:

[0070] Steady-state scaling geometric mean The calculation formula is:

[0071] Where P represents the effective value of active power, Q represents the effective value of reactive power, and H represents the harmonic content.

[0072] Long-period geometric mean The calculation formula is:

[0073] in, This represents the median power over a 24-hour moving window. Characterizing differences in class schedules, This represents the average value of harmonic content throughout the day.

[0074] The calculation formula is:

[0075] The calculation is as follows:

[0076] in, This indicates the harmonic content for each hour.

[0077] By fusing the geometric mean of transient scale, steady-state scale, and long-period scale across scales, the average geometric feature across scale time is obtained, expressed by the formula:

[0078] in, This represents the average value of the cross-scale time geometric features.

[0079] By combining the geometric mean of transient scale, the geometric mean of steady-state scale, the geometric mean of long-period scale, and the geometric mean of cross-scale time features, a multi-time-scale geometric mean feature set is constructed.

[0080] Geometric mean feature set and multi-timescale geometric mean feature set:

[0081]

[0082] The basic feature set BF and the geometric mean feature set and multi-timescale geometric mean feature set Perform fusion to obtain a fused feature set:

[0083] Since the calculation process primarily involves multiplication, it can reduce computer computational efficiency. Therefore, an optimized calculation formula is introduced, using logarithmic transformation to optimize the calculation of the geometric mean, converting multiplication into addition, thus simplifying the calculation process. Furthermore, the application of logarithmic transformation performs logarithmic processing on features at different scales separately, avoiding numerical overflow when multiplying across scales.

[0084] The basic principle of logarithmic transformation: For each basic electrical characteristic Their geometric mean GM is defined as:

[0085] Taking the natural logarithm of both sides of the formula, the formula becomes:

[0086] Using this method, determining the geometric mean is transformed into finding the arithmetic mean in logarithmic form. The calculation steps are as follows: 1. Logarithmic calculation First, for each basic electrical characteristic Find its natural logarithm This step transforms the original multiplication problem into an addition problem.

[0087] 2. Calculation of average value Next, the arithmetic mean of the resulting logarithms is calculated, expressed by the formula:

[0088] in, This represents the logarithm of the mean.

[0089] 3. Exponential Conversion Finally, by performing an exponential operation on the logarithm of the mean, the original geometric mean is obtained:

[0090] Logarithmic transformations play a crucial role in improving the stability of numerical computation, especially when dealing with extremely high or low values. Furthermore, addition and exponentiation are often faster than multiplication, which can significantly improve computational efficiency in parallel computing scenarios.

[0091] Step 3: Based on the fused feature set, train a long short-term memory network to process, identify, and distinguish the energy consumption patterns of each independent load in the industrial load.

[0092] 1) Acquire electrical signals and perform preprocessing, extract basic electrical features from the electrical signals, and construct a fusion feature set.

[0093] Preprocessing includes removing outliers and noise, while filling in or deleting missing data points.

[0094] 2) Then, the fused feature set data is standardized to a uniform range to accelerate training and improve model performance. The time series data is split into fixed-length windows for training the LSTM model.

[0095] The LSTM model takes a fused feature set as input and outputs the energy consumption of each load. Through its gating mechanism, it learns the dynamic temporal characteristics of industrial loads, establishing a mapping relationship from aggregated electrical energy to the energy consumption of individual equipment. The model takes the fused features as input, extracts load operating patterns through hidden layers, and outputs real-time energy consumption predictions for each independent device, achieving accurate decomposition of nonlinear loads such as frequency converters. This modeling process simultaneously optimizes network weights to ensure stable energy consumption identification capabilities even under industrial electromagnetic interference environments.

[0096] 3) Choose either the Adam or RMSprop optimizer to update the network weights. Set parameters such as learning rate, batch size, and training epoch. The data is propagated forward through the LSTM network to calculate the predicted output, while simultaneously comparing the predicted output with the actual energy consumption using a loss function. The gradient is calculated through backpropagation based on the loss value, and the network parameters are updated using the optimizer based on the gradient. Adjust the learning rate, batch size, or training epoch to fine-tune the parameters based on validation performance. Stop training when performance on the validation set no longer improves to avoid overfitting.

[0097] 4) Performance evaluation: Use test machine data to evaluate the predictive performance of the model and calculate performance indicators.

[0098] The trained LSTM model is validated in multiple dimensions using an independent test set. The overall deviation of power prediction is quantified by the root mean square error, and the measurement stability under steady-state conditions is evaluated by the mean absolute error. The energy decomposition accuracy index is introduced to comprehensively reflect the practical performance of the system in complex industrial scenarios. Finally, the model parameters are adjusted based on the test results.

[0099] Prepare the test set. Collect actual electricity consumption data from the same source as the training data, ensuring that this data is later in time than the training data to guarantee it wasn't used for model training. Perform the same preprocessing steps on these test data as on the training data, including outlier removal, handling missing values, and noise reduction. Ensure the quality and characteristics of the test data (e.g., sampling frequency, signal range) match the training data so the model can make predictions under similar conditions. Split the test data into sequences in the same format as the training data, including time window selection and feature extraction. Determine the size of the test set, allocating the majority of the data for training, a small portion for validation, and the remainder for testing. Before applying the test set, perform preliminary analyses, such as statistical analysis and visualization, to confirm the consistency and representativeness of the data.

[0100] Perform model predictions. Use the trained LSTM model to predict time series data in the test set. During the prediction process, check and handle any outliers or predictions that do not meet expectations, and record the model's prediction for each test sample.

[0101] Evaluate the overall performance of the model. Compare the results of different performance metrics to assess the overall performance of the model. Analyze the model's performance when handling different types of loads or data from different time periods. Create charts, such as comparison charts of predicted and actual values, and error analysis charts, to visually demonstrate the model's performance.

[0102] Step 4: Utilize the trained Long Short-Term Memory (LSTM) network model to achieve real-time energy consumption monitoring and decomposition of various similar loads in the industrial load.

[0103] The trained model outputs energy consumption predictions for each load. Simultaneously, it monitors for potential anomalies, data loss, and excessive bias in real time.

[0104] This invention solves the problem of energy consumption confusion among similar equipment by introducing harmonic content as a load fingerprint feature. The feature value considering harmonic content fully accounts for the impact of harmonics on industrial load equipment, improving equipment efficiency and lifespan while ensuring normal operation and thus improving the quality of machined products. It proposes a multi-timescale geometric average feature to simultaneously capture millisecond-level transients and hourly trends, resolving the contradictory requirements of industrial load monitoring that need to capture both millisecond-level transient details (such as harmonic abrupt changes during switching) and hourly operating trends. It achieves effective extraction and synergistic fusion of multi-timescale features from transient to long-period; and it uses logarithmic transformation to optimize geometric average calculation, converting multiplication into addition, improving computational efficiency by over 30% and avoiding numerical overflow.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An industrial load decomposition method based on the fusion of harmonic coupling and nonlinear characteristics, characterized in that, include: Collect electrical signals at the main power source of the industrial load; Based on the electrical signals, basic electrical features are extracted, and a basic feature set is constructed; Geometric mean features are determined by basic electrical characteristics, and a geometric mean feature set and a multi-time-scale geometric mean feature set are constructed. A fused feature set is constructed by combining the basic feature set, the multi-time-scale geometric mean feature set, and the geometric mean feature set. Train a long short-term memory network using a fused feature set; Non-invasive industrial load energy consumption monitoring is performed using a trained long short-term memory network.

2. The method according to claim 1, characterized in that, The electrical signals include: The electrical signals of the R-phase and Y-phase of the three-phase current signal.

3. The method according to claim 1, characterized in that, The basic feature set includes: Active power, reactive power, apparent power, power factor, harmonic content.

4. The method according to claim 1, characterized in that, The geometric mean feature set includes: The geometric mean of active power and reactive power, the geometric mean of reactive power and apparent power, the geometric mean of active power and current, the geometric mean of apparent power and current, the geometric mean of active power and harmonic content, the geometric mean of reactive power and harmonic content, the geometric mean of apparent power and harmonic content, the geometric mean of active power, reactive power and apparent power, and the geometric mean of active power, reactive power, apparent power and harmonic content.

5. The method according to claim 1, characterized in that, It also includes preprocessing the electrical signals: Remove outliers and noise, and fill in or delete missing data points.

6. The method according to claim 1, characterized in that, The multi-timescale geometric mean feature set includes: Transient-scale geometric mean, steady-state-scale geometric mean, long-period-scale geometric mean, and cross-scale time geometric mean.

7. The method according to claim 1, characterized in that, The method of training a long short-term memory network using a fused feature set includes: The Long Short-Term Memory (LSTM) network takes a fused feature set as input and outputs the energy consumption of each load.

8. The method according to claim 4, characterized in that, The geometric mean feature is calculated by logarithmic transformation, including: The formula for calculating the geometric mean is: in, This represents the basic electrical characteristics; n represents the total number of basic electrical characteristics involved in the calculation. Calculate the logarithmic mean: in, It represents the geometric mean.

Images