Non-invasive load decomposition method from multiple electrical parameters to points based on neural network

Abstract

The present invention relates to non-intrusive load decomposition, specifically to a non-intrusive load decomposition method from multiple electrical parameters to points based on a neural network. Various electrical parameters at the bus end are collected and optimized to obtain the most suitable optimized parameter group; the optimized parameter group is fitted in a way from multiple electrical parameters to points to obtain a fitting result; based on the corresponding mapping relationship between the optimized parameter group and the fitting result, the non-intrusive load decomposition model is trained to obtain a trained non-intrusive load decomposition model; the trained non-intrusive load decomposition model is used to decompose the various electrical parameters at the bus end collected in real time to obtain branch load data; the technical solution provided by the present invention can effectively overcome the defect that the prior art cannot perform non-intrusive load decomposition well due to the difficulty in selecting a suitable window size.

Description

Technical Field

[0001] The present invention relates to non-intrusive load disaggregation, and particularly to a non-intrusive load disaggregation method from multi-electrical parameters to points based on a neural network. Background Art

[0002] There is a large amount of energy waste in traditional power systems. In the context of the era of energy conservation and environmental protection, efficient management and effective utilization of electric energy have become the focus of research in the power field. In industrial production, non-intrusive load disaggregation technology is widely used, which is of great value for helping industrial users save energy and reduce production costs. Industrial equipment loads generate fewer events and need to be studied on a longer time scale. At present, the digital level of industrial users is not high, and large, medium and small industrial users cannot guarantee data quality while completing high-frequency industrial load data collection. At the same time, the current data collection, communication and storage capabilities cannot support high-frequency collection of industrial load data. Therefore, in practical applications, it is more practical to study ultra-low frequency load disaggregation in the industrial field.

[0003] Since Hart et al. proposed an algorithm called Transient Detection Algorithm in 1985, non-intrusive load disaggregation has gradually become a research hotspot. Mauch et al. implemented a deep recurrent network using long short-term memory (LSTM) neurons, which uses the aggregated values at the same time point to predict the energy consumption of the device at a specific time point. Based on the same principle, Thi-Thu-Huong et al. compared the performance of gated recurrent unit (GRU) and traditional recurrent network in the energy disaggregation task. In the experiment, GRU achieved better results. Zhang achieved state-of-the-art results through a deep convolutional network, which uses an aggregated data window to predict the midpoint value of the consumption window of the same device. They named the above method "sequence-to-point", which has the advantages of fast training speed, low model complexity, wide application range and effective avoidance of error accumulation.

[0004] However, in an ultra-low frequency industrial environment, the interval between each sampling of load data is longer and the load characteristics are more complex. For the existing sliding window methods, at a long window scale, the load characteristics will lose their meaning due to too long a time span, and at a short window scale, it is difficult to obtain sufficient feature quantities to express more complex load conditions. Summary of the Invention

[0005] (1) Technical Problems to be Solved

[0006] In view of the above-mentioned disadvantages of the prior art, the present invention provides a non-intrusive load decomposition method from multiple electrical parameters to points based on a neural network, which can effectively overcome the defect of the prior art that it is difficult to select a suitable window size and thus unable to perform non-intrusive load decomposition well.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions:

[0009] A non-intrusive load decomposition method from multiple electrical parameters to points based on a neural network, comprising the following steps:

[0010] S1. Collect various electrical parameters at the bus end and optimize them to obtain the most suitable optimized parameter group;

[0011] S2. Fit the optimized parameter group in a way of multiple electrical parameters to points to obtain a fitting result;

[0012] S3. Based on the corresponding mapping relationship between the optimized parameter group and the fitting result, train the non-intrusive load decomposition model to obtain a trained non-intrusive load decomposition model;

[0013] S4. Use the trained non-intrusive load decomposition model to decompose the various electrical parameters collected in real time at the bus end to obtain branch load data.

[0014] Preferably, in S1, collecting various electrical parameters at the bus end and optimizing them to obtain the most suitable optimized parameter group includes:

[0015] Collect various electrical parameters at the bus end, perform a correlation analysis with the parameters corresponding to the branch load decomposition, set different parameter groups according to the correlation analysis results, and select the most suitable optimized parameter group through experimental analysis;

[0016] Among them, the electrical parameters include total active power, total reactive power, total apparent power, and the current, voltage, admittance, active power, reactive power, and apparent power of each branch phase.

[0017] Preferably, in S2, fitting the optimized parameter group in a way of multiple electrical parameters to points to obtain a fitting result includes:

[0018] Without considering the window size, perform point-to-point fitting on the optimized parameter group. At this time, the mapping relationship is: the input is the optimized parameter group Y at time t at the bus end X1:XN , and the output is the electrical parameter x at time t of the branch τ .

[0019] Preferably, the mapping relationship is expressed by the following formula:

[0020]

[0021] Among them, L is the sequence length, and Y X1:XN is the optimal parameter group at bus terminal at time t, and Y X1:XN includes N electrical parameters X1, X2, …, XN, and θp is the network parameter.

[0022] Preferably, in S3, based on the corresponding mapping relationship between the optimal parameter group and the fitting result, the non-intrusive load decomposition model is trained to obtain a trained non-intrusive load decomposition model, including:

[0023] Based on the corresponding mapping relationship between the optimal parameter group and the fitting result, the deep neural network model is trained to obtain a trained deep neural network model.

[0024] Preferably, in S4, the trained non-intrusive load decomposition model is used to decompose various electrical parameters of the bus terminal collected in real time to obtain branch load data, including:

[0025] Collect various electrical parameters of the bus terminal in real time, and optimize the various electrical parameters of the bus terminal collected in real time to obtain the most suitable real-time optimal parameter group;

[0026] Use the trained deep neural network model to decompose the real-time optimal parameter group to obtain branch load data.

[0027] (III) Beneficial effects

[0028] Compared with the prior art, the non-intrusive load decomposition method from multi-electrical parameters to points based on neural network provided by the present invention proposes an industrial load decomposition Mep2point model from multi-electrical parameters to points, and uses the method of multi-electrical parameters to points to fit the optimal parameter group to obtain the load to be decomposed, without using a sliding window, which can effectively solve the problem that in an ultra-low frequency industrial environment, the sliding window type method cannot perform non-intrusive load decomposition well because it is difficult to select a suitable window size. Description of the drawings

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0030] Figure 1 is the flowchart of the present invention;

[0031] Figure 2 Schematic diagram of fitting the optimal parameter set by the method of multiple electrical parameters to points in the present invention;

[0032] Figure 3 Neural network architecture diagram of the non - invasive load decomposition model of the present invention;

[0033] Figure 4 Schematic diagrams of the prior art solutions of three sliding window classes of sequence - to - sequence, sequence - to - short - sequence, and sequence - to - point;

[0034] Figure 5 Line charts of evaluation indexes for load decomposition of three branch lines by sequence - to - short - sequence and sequence - to - point under different window sizes (11, 23, 55, 99, 199, 299);

[0035] Figure 6 Schematic diagrams of load decomposition results of Line1 by sequence - to - short - sequence and sequence - to - point under different window sizes (11, 23, 55, 99, 199, 299);

[0036] Figure 7 Schematic diagrams of load decomposition results of Line2 by sequence - to - short - sequence and sequence - to - point under different window sizes (11, 23, 55, 99, 199, 299);

[0037] Figure 8 Schematic diagrams of load decomposition results of Line3 by sequence - to - short - sequence and sequence - to - point under different window sizes (11, 23, 55, 99, 199, 299). Detailed implementation manners

[0038] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0039] The non - invasive load decomposition method based on neural network with multiple electrical parameters to points, as Figure 1 shown, ① Collect various electrical parameters at the bus end and optimize them to obtain the most suitable optimal parameter set, specifically including:

[0040] Collect various electrical parameters at the bus end, perform correlation analysis with the parameters corresponding to branch line load decomposition, set different parameter sets according to the correlation analysis results, and select the most suitable optimal parameter set through experimental analysis;

[0041] Among them, the electrical parameters include total active power, total reactive power, total apparent power, as well as the current, voltage, admittance, active power, reactive power, and apparent power of each branch phase.

[0042] ② As Figure 2 shown, the method of using multiple electrical parameters to points is adopted to fit the optimal parameter group, and the fitting result is obtained, specifically including:

[0043] Without considering the window size, the optimal parameter group is fitted point to point. At this time, the mapping relationship is: the input is the optimal parameter group Y at time t at the bus end X1:XN , and the output is the electrical parameter x at time t of the branch line τ .

[0044] Specifically, the mapping relationship is expressed by the following formula:

[0045]

[0046] Among them, L is the sequence length, Y X1:XN is the optimal parameter group at time t at the bus end, Y X1:XN includes N electrical parameters X1, X2,..., XN, and θp is the network parameter.

[0047] ③ Based on the corresponding mapping relationship between the optimal parameter group and the fitting result, the non-intrusive load decomposition model (as Figure 3 shown) is trained to obtain a trained non-intrusive load decomposition model, specifically including:

[0048] Based on the corresponding mapping relationship between the optimal parameter group and the fitting result, the deep neural network model is trained to obtain a trained deep neural network model.

[0049] ④ Using the trained non-intrusive load decomposition model to decompose the various electrical parameters of the bus end collected in real time, the branch line load data is obtained, specifically including:

[0050] Collect various electrical parameters of the bus end in real time, and optimize the various electrical parameters of the bus end collected in real time to obtain the most suitable real-time optimal parameter group;

[0051] Using the trained deep neural network model to decompose the real-time optimal parameter group to obtain the branch line load data.

[0052] Experimental analysis

[0053] 1) Dataset composition

[0054] Select an actual industrial load dataset for the experiment. This dataset contains the power data of a power distribution room in a large factory with a sampling frequency of 1 / 120 Hz from February 1, 2021 to April 30, 2021. Select 3 different types of branch lines: Line1, a steadily shutdown device: the main motor of the coal mill; Line2, a complex fluctuating device: the exhaust fan at the kiln head; Line3, a fluctuating startup device: the power supply of the electrical room at the kiln head. Collect relevant electrical parameters such as voltage, current, power, and admittance at the bus end, and collect the electrical parameters to be decomposed at each branch line, such as active power.

[0055] 2) Evaluation indicators

[0056] Select the Mean Absolute Error, Mean Square Error, Root Mean Square Error, and correlation coefficient CC to evaluate the load decomposition results of each load decomposition model.

[0057] Among them, MSE is the average of the squares of the differences between the predicted values and the true values. Compared with MAE, MSE can retain the positive and negative information of the errors and pay more attention to the influence of large errors, because after the errors are squared, large errors will get greater weights. RMSE is the square root of MSE, providing a measurement standard similar to the target variable, making it more intuitive to compare and interpret the results. And MAE is the average of the absolute values of the differences between the predicted values and the true values, which measures the average size of the prediction errors.

[0058] 3) Experimental settings

[0059] Using the actual industrial load dataset, compare the decomposition effects of the Mep2point model for industrial load decomposition from multiple electrical parameters to points in the technical solution of this application with two sliding window-based load decomposition models Seq2point and Seq2subseq in the non-intrusive load decomposition field (as Figure 4 shown) for the load of 3 different types of branch lines. The comparison results are shown in the following table:

[0060] Table 1 Decomposition effect table of different load decomposition models

[0061]

[0062] Table 2 Comparison table of decomposition effects between different load decomposition models

[0063]

[0064] In Line1, the Mep2point model of this application has good fitting results for all 5 downward trends. There are only two very short instantaneous upward trends in the 5th downward trend, but this will not interfere with the load decomposition results. However, there are large errors in the other two sliding window-based load decomposition models. Seq2point has more error fluctuations, while Seq2subseq cannot accurately capture the 0 point. The emergence of the above problems is determined by the nature of these two models. Seq2point fits a single point, so there will be greater perturbations; Seq2subseq fits a sequence, which can better fit the general curve but is lacking in details.

[0065] In Line2, it can be clearly seen the superiority of the Mep2point model of this application, which has a better fitting curve and better evaluation indicators, while the other two sliding window-based load decomposition models are very ineffective at this time. In practical applications, the sliding window-based methods cannot obtain the actual operating conditions of such industrial equipment, while the Mep2point model of this application has a better load decomposition effect.

[0066] In the example of Line3, ten upward trends are intercepted. Seq2point and Seq2subseq mostly do not reach the peak value in the ten upward trends and cannot decompose the actual peak value, while the Mep2point model of this application can accurately decompose the maximum value of the load. However, the fluctuations of the load decomposition results are more complex at this time, resulting in a decrease in the evaluation indicators because Line3 is not an industrial equipment but the total power supply of a room, and there is greater interference at this time. Moreover, this application selects a series of bus terminal electrical parameters, which multiplies the interference, resulting in large fluctuations in the load decomposition results. The fitting curve can be smoothed through post-processing noise reduction.

[0067] The above experiments verified the superiority of the Mep2point model of this application in the actual industrial load dataset. Compared with the existing sliding window-based load decomposition models, the Mep2point model of this application can better obtain the actual operating conditions of industrial loads, so that it can more accurately monitor the load operating conditions in practical applications, which is helpful for the monitoring and analysis of industrial equipment.

[0068] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements will not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-invasive load decomposition method from multiple electrical parameters to points based on a neural network, characterized in that: Including the following steps: S1. Collect various electrical parameters at the bus end and optimize them to obtain the most suitable optimized parameter set; S2. Fit the optimized parameter set by the method of multiple electrical parameters to points to obtain a fitting result; S3. Based on the corresponding mapping relationship between the optimized parameter set and the fitting result, perform model training on the non-intrusive load decomposition model to obtain a trained non-intrusive load decomposition model; S4. Use the trained non-intrusive load decomposition model to perform load decomposition on various electrical parameters collected in real time at the bus end to obtain branch load data; In S2, the optimized parameter set is fitted by the method of multiple electrical parameters to points to obtain a fitting result, including: Without considering the window size, the optimal parameter group is fitted point-to-point. The mapping relationship at this time is: the input is the optimal parameter group Y at the bus terminal at time t X1:XN , the output is the electrical parameter x of the branch line at time t τ ; The mapping relationship is expressed by the following formula: where L is the sequence length, and Y X1:XN is the preferred parameter set at bus end at time t, and Y X1:XN includes N electrical parameters X1, X2, …, XN, and θp is the network parameter; In S3, based on the corresponding mapping relationship between the optimized parameter set and the fitting result, perform model training on the non-intrusive load decomposition model to obtain a trained non-intrusive load decomposition model, including: Based on the corresponding mapping relationship between the optimized parameter set and the fitting result, perform model training on the deep neural network model to obtain a trained deep neural network model.

2. The non-intrusive load decomposition method from multiple electrical parameters to points based on a neural network according to claim 1, characterized in that: In S1, various electrical parameters at the bus end are collected and optimized to obtain the most suitable optimized parameter set, including: Collect various electrical parameters at the bus end, perform correlation analysis with the parameters corresponding to branch load decomposition, set different parameter sets according to the correlation analysis results, and select the most suitable optimized parameter set through experimental analysis; Among them, the electrical parameters include total active power, total reactive power, total apparent power, and the current, voltage, admittance, active power, reactive power, and apparent power of each phase.

3. The non-invasive load decomposition method from multiple electrical parameters to points based on a neural network according to claim 1, characterized in that: In S4, use the trained non-intrusive load decomposition model to perform load decomposition on various electrical parameters collected in real time at the bus end to obtain branch load data, including: Collect various electrical parameters at the bus end in real time, and optimize the various electrical parameters collected in real time at the bus end to obtain the most suitable real-time optimized parameter set; Use the trained deep neural network model to perform load decomposition on the real-time optimized parameter set to obtain branch load data.

Images