A non-intrusive load identification method based on user prior information and XGBoost-LSTM

Abstract

The present application belongs to the field of commercial user power load electricity monitoring, and particularly relates to a non-intrusive load identification method based on user prior information and XGBoost-LSTM. A baseline load model is established by using XGBoost to effectively eliminate the relatively stable non-air conditioning load component in the total load; residual load sequence is modeled by combining LSTM, and user behavior prior information such as business hours, holidays and environmental characteristics such as meteorological temperature are used to capture the dynamic law of central air conditioning load changing with time and environment; through joint optimization of XGBoost and LSTM models, load decomposition of total load into baseline load and central air conditioning is realized. The method can maintain high identification accuracy under low sampling rate data, and provide reliable data support for commercial building central air conditioning energy consumption decomposition and energy efficiency management.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of commercial user power load online monitoring and intelligent analysis, and particularly relates to a non-intrusive load recognition method based on user prior information and XGBoost-LSTM. BACKGROUND

[0002] Load monitoring, as a key process of smart grid technology research, includes intrusive load monitoring and non-intrusive load monitoring. The intrusive load monitoring technology needs to install a power consumption monitoring device on each power load, resulting in complex installation and deployment and high cost. The non-intrusive load monitoring technology (NILM) only needs to install a monitoring terminal device at the power supply inlet of the monitoring system, so as to realize online monitoring of different load devices. Compared with the traditional intrusive method, NILM is the main direction of future development, can realize fine management of the demand side, has strong information security, and has low operation and maintenance cost.

[0003] In a commercial scenario, the central air conditioning load is usually closely related to business hours, meteorological temperature, holidays and other factors, and the operation law thereof has obvious time dependence and environmental sensitivity.

[0004] The existing non-intrusive load monitoring method mainly relies on the transient characteristics of high sampling rate current and voltage signals to identify the start-stop state of household appliances. However, in a commercial scenario, the smart meter generally only provides power data at a minute level or lower frequency, and lacks the high-frequency characteristics required by the traditional method, resulting in a significant decrease in recognition accuracy. At the same time, since the load curve is often disturbed by the mixed operation of multiple devices, simply relying on power time series mode for decomposition is easy to cause misjudgment and deviation.

[0005] Chinese patent application CN 119829980 A discloses a method suitable for non-interventional identification of central air conditioner operating load, selects central air conditioners of different brands, records the electrical waveform of the complete operation process, performs time-frequency domain feature analysis, extracts common features, selects the total daily electricity load data and daily air temperature data of the user in the past year, analyzes the correlation between the user's daily electricity load curve and the daily air temperature curve, obtains the heating load area, the basic load area and the cooling load area, and then performs linear data fitting with temperature to obtain the load curve. Subtract the cooling load curve and the basic load curve in it to obtain the relationship formula between the steady-state power of the user's central air conditioner and the temperature, and further judge whether the central air conditioner is in the start / stop state. During the start event and the stop event of the user's air conditioner, the sub-item electricity is estimated, and the running electricity of the central air conditioner equipment during operation is obtained by integrating the steady-state power, so as to realize the running state tracking and electricity estimation of the central air conditioner. The deficiency of the method is that the smart meter generally only provides minute-level or lower frequency electricity data, resulting in a significant decrease in recognition accuracy. At the same time, the load curve is often disturbed by the mixed operation of multiple devices.

[0006] Chinese patent application CN 118839806 A discloses a multi-factor load prediction method and device based on LSTM-XGBoost neural network, which comprises: collecting source data, generating waiting analysis data based on historical load data and historical external factor data after preprocessing; generating a plurality of first input data sets according to the data type of the waiting analysis data, and extracting the long-term dependence relationship in each input data set according to the constructed double-layer LSTM model; generating a second input data set based on the long-term dependence relationship and the historical load data, fitting and predicting the second input data set according to the constructed XGBoost model, and generating a multi-factor load prediction value based on the LSTM-XGBoost combined model. The application first uses the LSTM neural network to extract the long-term dependence relationship of the historical data, then uses the XGBoost model for fitting, and finally completes the prediction, thereby improving the accuracy of load prediction. The LSTM-XGBoost combined model used by the application uses the XGBoost model to fit the historical load data and the long-term dependence relationship extracted by the LSTM neural network, combines the advantages of the two single models, obtains a more accurate and more applicable prediction model. The deficiency of the method is that the file only provides a mathematical model, does not provide a real use case, and does not provide a technical means combined with real working conditions. SUMMARY

[0007] In view of the problem of insufficient load identification accuracy under low sampling rate smart meter data, the present application provides a non-intrusive load identification method based on user prior information and XGBoost-LSTM, which improves the load identification accuracy.

[0008] The application is implemented by a non-intrusive load identification method based on user prior information and XGBoost-LSTM, which comprises the following steps:

[0009] Step S1, data preprocessing and feature construction: the minute-level total load time series collected by the intelligent electric meter in the commercial scenario are preprocessed, and on this basis, the input feature matrix is constructed by combining the user prior information and the time sequence characteristics of the load curve;

[0010] Step S2, XGBoost baseline load modeling: the total load sequence is modeled by using the XGBoost gradient boosting tree, the baseline load estimate value is fitted, and the air conditioner dominant component in the total load is separated;

[0011] Step S3, LSTM residual load modeling: the residual sequence between the total load and the baseline load estimate value is calculated, which is used as the candidate signal of the central air conditioner load, the residual sequence is modeled by using the LSTM, and the dynamic time sequence mode of the central air conditioner load is extracted by combining the time sequence characteristics and the user prior information; the LSTM can capture the long-term dependence characteristics across time, and overcome the defects of the traditional model in identifying the air conditioner start-stop and load change response under the condition of low sampling rate.

[0012] Step S4, fusion of XGBoost baseline load model and LSTM residual load model: in the model output stage, the baseline load provided by the XGBoost and the central air conditioner load provided by the LSTM are added together to reconstruct the total load sequence. By jointly optimizing the loss function, it is ensured that the estimate value has high fitting accuracy on both the overall load and the sub-load levels. The final output result includes the baseline load estimate value and the central air conditioner load estimate value.

[0013] The baseline load model is established by using the XGBoost, which effectively eliminates the relatively stable non-air conditioner load component in the total load; the residual load sequence is modeled by using the LSTM, the user behavior prior information such as business hours, holidays and the environmental characteristics such as meteorological temperature are used to capture the dynamic law of the central air conditioner load changing with time and environment; by jointly optimizing the XGBoost and LSTM models, the load decomposition of the total load into the baseline load and the central air conditioner load is realized. The method can maintain high identification accuracy under low sampling rate data, and provide reliable data support for commercial building central air conditioner energy consumption decomposition and energy efficiency management.

[0014] As a preferred scheme:

[0015] The user prior information in step S1 includes business hours, holiday markers and temperature.

[0016] The time sequence characteristics of the load curve in step S1 include historical load, sliding mean and load change rate.

[0017] The preprocessing of step S1 includes cleaning, interpolation and format unification, and removing obvious outliers and missing data.

[0018] The preprocessing includes timestamp verification and alignment, missing value interpolation, outlier removal and short-term smoothing processing, and the total load sequence is For short-term missing less than 30 minutes, linear interpolation is used: , where k is the index of consecutive missing points; for medium-term missing of 30-360 minutes, the historical daily average value of the same time period is filled: , where is the load at the time point on the dth day in history, and N is the number of historical days; for long-term missing greater than 360 minutes, it is marked as unavailable data and excluded from model training;

[0019] The outlier detection uses the Hampel method based on the local median in the window W and the absolute median difference If , it is determined to be abnormal, and the abnormal point is replaced by the window median ;

[0020] The short-term smoothing processing refers to the exponential weighted moving average processing of the sequence.

[0021] The feature matrix of step S1 is divided into baseline features x t and air conditioning environment coupling features z t in the construction phase,

[0022] The baseline features x t describe the periodicity of the load and the user's electricity consumption behavior, including hour sine / cosine encoding: , where is the hour index; historical moving average and difference features: , used to describe the short-term trend and change rate of the load; holiday and business hours markers as binary features: ;

[0023] The air conditioning environment coupling features z t reflect the dynamic influence of meteorological conditions on air conditioning load, based on temperature data from weather forecasts, an hour-level temperature sequence is constructed using sine fitting method: , where h t is the hour index, h min is the hour when the minimum temperature occurs, set to 2:00; based on this hour-level sequence, rolling average temperature and temperature change rate can be constructed: and cooling degree days (CDD / HDD): where T base is the baseline temperature, set as 22℃; meanwhile, the interaction feature between temperature and business hours is constructed: .

[0024] The continuous features of the feature matrix are robustly normalized based on median and interquartile range before inputting into the model, and the categorical features are integer encoded to ensure the consistency of feature scales and the stability of model training.

[0025] The XGBoost baseline load modeling of step S2 includes the following steps:

[0026] Step S21, input and output definition: the input is the preprocessed feature matrix Xt, containing baseline features x t : time period features such as hour sine / cosine encoding, holiday label and business hours label, and historical load statistics such as sliding mean, sliding standard deviation and difference features, and the model output is the baseline estimate of total load , which is expressed as:

[0027] ,

[0028] where f XGB (·) represents the XGBoost model, and Θ is the tree structure parameter and model hyperparameter;

[0029] Step S22, model training: based on the minute-level total load data, the XGBoost model is trained using historical data, and the model is fitted by the iterative addition combination of multiple regression trees. Each tree minimizes the objective function during training:

[0030] ,

[0031] where the first term is the mean square error loss function, which measures the difference between the prediction and the true total load; the second term is the regularization term Ω(f k ), which is used to control the complexity of the tree to prevent overfitting. Each tree is constructed by the gradient boosting algorithm, using the residual of the previous round as the fitting target of the next tree to realize the step-by-step approximation of the baseline load;

[0032] Step S23, hyperparameter setting and cross-validation: to improve the accuracy and stability of baseline load modeling, cross-validation is used to adjust key hyperparameters, including the maximum depth of the tree max_depth, the learning rate eta, the subsample ratio subsample and the colsample_bytree ratio, by minimizing the MSE on the validation set, the optimal parameter combination is selected;

[0033] Step S24, residual sequence construction: after the baseline load estimation is completed, the difference between the baseline load and the actual total load is taken as the residual sequence: , which provides an input signal for LSTM time series modeling.

[0034] The step S3, the LSTM residual load modeling, comprises the following steps:

[0035] Step S31, input data construction: the model input is in the form of a sliding window of the residual sequence, and the environmental coupling feature z is combined t : temperature rolling average and temperature change rate; CDD / HDD refrigeration / heating metric, business hours and temperature interaction feature, which is specifically represented as:

[0036] , wherein L is the length of the sliding window, and is set to 2 hours;

[0037] Step S32, network structure and training strategy: the LSTM model output is the central air conditioner load estimation value at the current time step :

[0038] ,

[0039] , wherein fLSTM(·) represents the LSTM forward calculation function, and Φ is the network weight and bias parameter; the model training adopts a mean square error loss function:

[0040] ,

[0041] , wherein N is the total number of samples, and the weight parameters are optimized through the back propagation and gradient descent algorithm;

[0042] Step S33, output and residual analysis: after the LSTM model training is completed, the central air conditioner load sequence is output.

[0043] The advantages and positive effects of the present application are that: firstly, the gradient boosting tree (XGBoost) is used to model the baseline load, and the relatively stable electricity consumption component is removed from the total load; then, the long short-term memory network (LSTM) is used to model the residual load and environmental features, and the central air conditioner load component is extracted; finally, the accurate decomposition of the total load is realized through joint optimization, and the problem of insufficient applicability of traditional NILM methods in commercial scenarios is effectively solved.

[0044] The present application can capture long-term dependence features across time, and overcome the defects of traditional models in identifying air conditioner start-stop and load change response under low sampling rate conditions.

[0045] The application faces the non-intrusive load identification problem of commercial scenario central air conditioning, and provides a load identification method combining user prior information and XGBoost-LSTM. The method can accurately decompose the central air conditioning load curve under the conditions of noise interference in total load data, low sampling rate limitation and coupling of multiple loads. Compared with traditional load identification methods, the application has stronger robustness and generalization ability, and can be widely used in load decomposition and energy efficiency analysis requirements under different building types and different climate conditions. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a step block diagram of the application.

[0047] Figure 2 is a principle block diagram of the application. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the application more clear, the application will be further described in detail below combined with examples.

[0049] Example 1

[0050] As shown in Figures 1-2 , the non-intrusive load identification method based on user prior information and XGBoost-LSTM of the application adopts the following technical scheme:

[0051] Step S1, data preprocessing and feature construction: the minute-level total load time series collected by the intelligent electric meter in the commercial scenario is preprocessed, and on this basis, the input feature matrix is constructed combining the user prior information and the time sequence characteristics of the load curve;

[0052] Step S2, XGBoost baseline load modeling: the total load sequence is modeled by using XGBoost gradient boosting tree, the baseline load estimation value is fitted, and the air conditioner dominant component in the total load is separated;

[0053] Step S3, LSTM residual load modeling: the residual sequence between the total load and the baseline load estimation value is calculated, which is used as the candidate signal of the central air conditioning load, the residual sequence is modeled by using LSTM, the dynamic time sequence mode of the central air conditioning load is extracted combining the time sequence characteristics and the user prior information; LSTM can capture the long-term dependence characteristics across time, and overcome the defects of traditional models in identifying the start and stop of air conditioner and the response of load change under the condition of low sampling rate.

[0054] Step S4: Fusion of the XGBoost baseline load model and the LSTM residual load model: In the model output stage, the baseline load provided by XGBoost and the central air conditioning load provided by LSTM are summed to reconstruct the total load sequence. By jointly optimizing the loss function, high fitting accuracy is ensured at both the overall load and sub-load levels. The final output includes: baseline load estimate and central air conditioning load estimate.

[0055] Example 2

[0056] Based on Example 1, the steps are further refined:

[0057] First, the historical total load data collected by smart meters throughout the year for a large commercial complex was obtained, with a sampling interval of 1 minute. Business hours, holiday markers, and corresponding meteorological temperature data were also collected. The load time series data underwent cleaning and preprocessing, specifically including timestamp verification and alignment, missing value imputation, outlier removal, and short-term smoothing.

[0058] Let the total load sequence be .

[0059] For short-duration missing data (≤30 minutes), linear interpolation is used: , where k is the index of consecutive missing points.

[0060] Moderate missing values ​​(30–360 minutes) are filled using the daily average of the same historical time period: ,in Let N be the load at this point in time on the d-th day of the history, and N be the number of historical days.

[0061] Data missing for extended periods (>360 minutes) is marked as unusable and excluded from model training.

[0062] Outlier detection uses a local median within a window W. With absolute median The Hampel method, if If it is not found, it is considered an anomaly, and the anomaly point is replaced with the window value. .

[0063] To further suppress high-frequency noise, the sequence is processed by exponentially weighted moving average (EWMA) while retaining the original signal for subsequent modeling.

[0064] During the feature construction phase, the input features are divided into baseline features x t Coupling characteristics with air-conditioned environment z t Two categories.

[0065] Baseline feature x tIt primarily depicts the periodicity of the load and user electricity consumption behavior, including hourly sine / cosine encoding: ,in Hourly indicators; historical moving average and difference characteristics: Used to characterize the short-term trend and rate of change of load; holiday markers and business hours markers are used as binary features: .

[0066] Air conditioning environment coupling characteristics z t This reflects the dynamic impact of meteorological conditions on air conditioning load. Based on temperature data from weather forecasts, an hourly temperature series is constructed using a sine fitting method. , where h t For hourly indexing, h min The hour at which the lowest temperature occurs is set to 2:00. Based on this hourly series, a rolling average temperature and the rate of temperature change can be constructed: And daily cooling / heating measurements (CDD / HDD): Among them, T base The reference temperature is set at 22℃.

[0067] At the same time, the interaction features between temperature and business hours were constructed: .

[0068] All continuous features are robustly normalized based on median and interquartile range (IQR) before being input into the model, while categorical features are encoded with integers to ensure feature scale consistency and model training stability.

[0069] Secondly, a gradient boosting tree (XGBoost) is used to perform baseline modeling of the total load time series in a commercial scenario. The components of the total load dominated by periodic patterns and user behavior are extracted, while the dynamic changes in air conditioning load are removed. The modeling process is as follows:

[0070] 1. Input and Output Definitions: The input is the preprocessed feature matrix Xt, containing the baseline features x. t Features include: time period characteristics (hourly sine / cosine encoding), holiday markings, business hours markings, and historical load statistics (moving mean, moving standard deviation, difference characteristics). The model output is a baseline estimate of the total load. The formula is expressed as:

[0071] ,

[0072] Where f XGB (·) represents the XGBoost model, and Θ represents the tree structure parameters and model hyperparameters.

[0073] 2. Model Training: Based on minute-level total load data, the XGBoost model is trained using historical data. The model fits the data through iterative additive combinations of multiple regression trees, with each tree minimizing the objective function during training.

[0074] ,

[0075] The first term is the mean squared error (MSE) loss function, which measures the difference between the predicted and actual total load; the second term is the regularization term Ω(f k This is used to control the complexity of the tree and prevent overfitting. Each tree is constructed incrementally using the gradient boosting algorithm, with the residual from the previous round used as the fitting target for the next tree, thus progressively approximating the baseline load.

[0076] 3. Hyperparameter Setting and Cross-Validation: To improve the accuracy and stability of baseline load modeling, cross-validation is used to adjust key hyperparameters, including the maximum tree depth (max_depth), learning rate (eta), subsample ratio (subsample), and column sampling ratio (colsample_bytree). The optimal parameter combination is selected by minimizing the mean squared error (MSE) on the validation set.

[0077] 4. Residual Sequence Construction: After the baseline load estimation is completed, the difference between the baseline load and the actual total load is used as the residual sequence. This provides input signals for LSTM timing modeling.

[0078] Finally, a Long Short-Time Memory (LSTM) network is used to dynamically model the residual sequence to identify the temporal characteristics of the central air conditioning load. The process is as follows:

[0079] 1. Input Data Construction: The model input is a sliding window of residual sequences, combined with environmental coupling features z. t Temperature rolling average and temperature change rate; CDD / HDD cooling / heating metrics, operating hours and temperature interaction characteristics. Specifically, this is expressed as:

[0080] ,

[0081] Where L is the sliding window length, set to 2 hours, to ensure that the model can learn short-term fluctuations and long-term trends.

[0082] 2. Network Structure and Training Strategy: The LSTM model output is the estimated central air conditioning load at the current time step. :

[0083] ,

[0084] Where fLSTM(·) represents the forward computation function of LSTM, and Φ represents the network weights and bias parameters. The model training uses the mean squared error loss function:

[0085] ,

[0086] Where N is the total number of samples. The weight parameters are optimized using backpropagation and gradient descent algorithms.

[0087] 3. Output and Residual Analysis: After the LSTM model is trained, the output is the central air conditioning load sequence. .

[0088] Finally, the baseline load estimate is combined with the dynamic residual load estimate to achieve a fine-grained decomposition of the total load. Baseline Load The residual loading is provided by the XGBoost model and output by the LSTM model. Supports the dynamic timing characteristics of central air conditioning.

[0089] The fusion process achieves total load reconfiguration through addition:

[0090] ,

[0091] in, This represents the total load after reconstruction. To improve fusion accuracy, a joint loss function is designed, optimizing both the total load and sub-load levels simultaneously.

[0092] ,

[0093] Here, α is the weighting coefficient between the residual load and the total load error. By minimizing this joint loss function, the model is ensured to both fit the overall load curve and accurately identify dynamic changes in the central air conditioning load. The final output includes baseline load estimates. Central air conditioning load estimate and reconfiguration of total load This enables non-intrusive load identification and energy consumption decomposition in commercial scenarios.

[0094] Experimental verification

[0095] Based on the above steps, load identification was carried out on the data of a commercial user from June 23 to July 8, 2025, and the start-stop information and power consumption results were compared and analyzed with the data of branch monitoring. The comparison is shown in the table below.

[0096] During air conditioner operation, the power consumption measured by the guide rail meter was 159695 kWh, and the power consumption estimated by the algorithm was 164834 kWh, with a power consumption estimation deviation of 4.4%.

[0097] Table 1. Verification Results of Air Conditioning Load Identification

[0098] Date Shunt monitored energy (kWh) Load identified energy (kWh) Identified / actual energy (%) June 23 12825 11985 93 June 25 4660 4750 102 June 24 5696 6498 114 June 26 4623 4764 103 June 27 4390 4979 113 June 28 4533 4368 96 June 29 5337 5472 103 June 30 10010 11199 112 July 1 9410 9938 106 July 2 10990 11817 108 July 3 14133 14603 103 July 4 5148 5986 116 July 5 4590 4591 100 July 6 5298 5572 105 July 7 5981 5671 95 July 8 5047 5437 108 .

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-intrusive load identification method based on user prior information and XGBoost-LSTM, characterized in that: Includes the following steps: Step S1, Data Preprocessing and Feature Construction: The minute-level total load time series collected by smart meters in commercial scenarios is preprocessed. Based on this, the input feature matrix is ​​constructed by combining user prior information and the time series features of the load curve. Step S2, XGBoost Baseline Load Modeling: The total load sequence is modeled using XGBoost gradient boosting tree to fit the baseline load estimate and separate the air conditioning-dominated component in the total load; Step S3, LSTM residual load modeling: Calculate the residual sequence between the total load and the baseline load estimate, use it as a candidate signal for the central air conditioning load, use LSTM to model the residual sequence, and combine time series characteristics and user prior information to extract the dynamic time series pattern of the central air conditioning load. Step S4, Fusion of XGBoost Baseline Load Model and LSTM Residual Load Model: In the model output stage, the baseline load provided by XGBoost and the central air conditioning load provided by LSTM are summed to reconstruct the total load sequence.

2. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 1, characterized in that, The prior information of the user in step S1 includes business hours, holiday markers, and temperature.

3. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 1, characterized in that, The time-series characteristics of the load curve in step S1 include historical load, moving average, and load change rate.

4. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 1, characterized in that, The preprocessing in step S1 includes cleaning, interpolation, and format unification, as well as removing outliers and missing data.

5. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 1, characterized in that, The steps also include step 5, which ensures that the estimated value has fitting accuracy at both the overall load and sub-load levels by jointly optimizing the loss function. The final output includes the baseline load estimate and the central air conditioning load estimate.

6. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 4, characterized in that, The preprocessing includes timestamp verification and alignment, missing value imputation, outlier removal, and short-run smoothing. Let the total load sequence be... For short-duration missing data (less than 30 minutes), linear interpolation is used. Where k is the index of consecutive missing points; for medium-duration missing points of 30–360 minutes, the average daily value of the same historical time period is used for filling: ,in The load at this time point on day d in history is N, where N is the number of historical days; data missing for a long period of more than 360 minutes is marked as unusable and excluded from model training; The outlier removal method uses the local median within a window W. With absolute median The Hampel method, if , If it is not found, it is considered an anomaly, and the anomaly point is replaced with the window value. ; The aforementioned short-term smoothing process refers to applying an exponentially weighted moving average to the sequence.

7. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 1, characterized in that, In the feature matrix construction phase of step S1, the input features are divided into baseline features x. t Coupling characteristics with air-conditioned environment z t Two categories, The baseline feature x t Characterizing the periodicity of load and user electricity consumption behavior, including hourly sine / cosine encoding: ,in Hourly indicators; historical moving average and difference characteristics: It is used to characterize the short-term trend and rate of change of load; Holiday markers and business hours markers are used as binary features: ; The air conditioning environment coupling characteristics z t This reflects the dynamic impact of meteorological conditions on air conditioning load. Based on temperature data from weather forecasts, an hourly temperature series is constructed using a sine fitting method. , where h t For hourly indexing, h min The hour at which the lowest temperature occurs is set to 2:00; based on this hourly series, a rolling average temperature and the rate of temperature change can be constructed: And daily cooling / heating measurements, i.e., CDD / HDD: T base The baseline temperature was set at 22℃; simultaneously, the interaction characteristics between temperature and business hours were constructed: .

8. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 7, characterized in that, The continuous features of the feature matrix are subjected to robust normalization based on median and interquartile range before being input into the model, and the categorical features are encoded with integers.

9. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 1, characterized in that, The XGBoost baseline load modeling in step S2 includes the following steps: Step S21, Input and Output Definition: The input is the preprocessed feature matrix Xt, containing the baseline features x t The model incorporates time-period features, including hourly sine / cosine coding, holiday markings, and business hour markings, as well as historical load statistics, including the moving mean, moving standard deviation, and difference characteristics. The model output is a baseline estimate of the total load. The formula is expressed as: ， Where f XGB (·) represents the XGBoost model, and Θ represents the tree structure parameters and model hyperparameters; Step S22, Model Training: Based on minute-level total load data, the XGBoost model is trained using historical data. The model is fitted by iterative addition of multiple regression trees, with each tree minimizing the objective function during training. ， The first term is the mean squared error loss function, which measures the difference between the predicted and the actual total load; the second term is the regularization term Ω(f k This is used to control the complexity of the tree and prevent overfitting. Each tree is built step by step through the gradient boosting algorithm, and the residual of the previous round is used as the fitting target of the next tree to gradually approach the baseline load. Step S23, Hyperparameter setting and cross-validation: To improve the accuracy and stability of baseline load modeling, cross-validation is used to adjust key hyperparameters, including the maximum tree depth (max_depth), learning rate (eta), subsample ratio (subsample), and column sampling ratio (colsample_bytree). The optimal parameter combination is selected by minimizing the MSE on the validation set. Step S24, Residual Sequence Construction: After the baseline load estimation is completed, the difference between the baseline load and the actual total load is used as the residual sequence. This provides input signals for LSTM timing modeling.

10. The non-intrusive load identification method based on user prior information and XGBoost-LSTM as described in claim 1, characterized in that, Step S3, LSTM residual load modeling, includes the following steps: Step S31, Input Data Construction: The model input is a sliding window of residual sequences, combined with environmental coupling features z. t Temperature rolling average and temperature change rate; CDD / HDD cooling / heating metrics, operating hours and temperature interaction characteristics, specifically expressed as: ,in Let L be the residual load sequence window of length L, and let L be 2 hours. Step S32, Network Structure and Training Strategy: The LSTM model output is the estimated central air conditioning load at the current time step. : ， in This represents the forward computation function of the LSTM, where Φ represents the network weights and bias parameters. The model training uses the mean squared error loss function. ， Where N is the total number of samples. Given a residual sequence, the weight parameters are optimized using backpropagation and gradient descent algorithms. Step S33, Output and Residual Analysis: After the LSTM model training is completed, the central air conditioning load sequence is output. .

Images