Bus peak load prediction method fusing transfer learning and peak load adaptive identification

Abstract

The invention discloses a bus peak load prediction method fusing transfer learning and peak load adaptive identification, and the method comprises the steps: constructing a dynamic weight-driven transfer learning framework, and achieving the precise migration of multi-source bus knowledge to a target bus; introducing a self-adaptive threshold mechanism based on load volatility and a peak value proportion, and dynamically identifying peak values and conventional load time periods; and designing a customized loss function for strengthening the peak error weight and using the customized loss function for fine adjustment of the model, thereby remarkably improving the load prediction precision of the bus in a key period. The target bus prediction precision is obviously improved through multi-source bus transfer learning; a self-adaptive threshold identification mechanism realizes dynamic judgment of a peak time period sensitive to load fluctuation, and the perception capability of the model to a key load time period is enhanced; through the synergistic effect of the technical means, the overall prediction accuracy is finally improved, meanwhile, the prediction precision and stability of the peak load are remarkably improved, and the actual requirements of a power grid for high precision and high robustness of load prediction are met.

Description

Technical Field

[0001] The present invention relates to the field of power system load forecasting, and in particular to a bus peak load forecasting method integrating transfer learning and peak load adaptive identification. Background Art

[0002] Existing load forecasting methods that model a single bus individually still lack accuracy in handling the complex spatiotemporal correlations between multiple buses within a region and the volatility of instantaneous peak loads. This makes it difficult to support the grid's scheduling and supply guarantee needs during high-load periods. The methods also have the following drawbacks:

[0003] (1) It is difficult to capture the temporal and spatial correlation between multiple bus loads in the same area, which limits the improvement of prediction accuracy;

[0004] (2) The prediction effect of busbar instantaneous peak load is poor.

[0005] With the development of data-driven technology, methods such as transfer learning, multi-task modeling and dynamic feature weighting have shown significant advantages in improving model generalization capabilities and prediction performance during critical periods, prompting load forecasting to evolve towards high resolution and high robustness. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the existing technology and provide a bus peak load forecasting method that integrates transfer learning and peak load adaptive identification. It innovatively constructs a dynamic weight-driven transfer learning framework to achieve accurate migration of multi-source bus knowledge to the target bus; introduces an adaptive threshold mechanism based on load volatility and peak share to dynamically identify peak and regular load periods; further designs a customized loss function that strengthens the peak error weight and uses it to fine-tune the prediction model, thereby significantly improving the load forecasting accuracy of the bus during critical periods.

[0007] The technical solution to achieve the above object is: a bus peak load prediction method integrating transfer learning and peak load adaptive identification, comprising the following steps:

[0008] S1: Build a dynamic weight-driven transfer learning framework to accurately transfer multi-source bus knowledge to the target bus through weight updates and model training of the prediction model;

[0009] S2, introduces an adaptive threshold mechanism based on load volatility and peak proportion to dynamically identify peak and regular load periods;

[0010] S3, designs a customized loss function that strengthens the peak error weight and uses it for model fine-tuning to improve the load forecasting accuracy of the bus during critical periods.

[0011] In the above-mentioned bus peak load forecasting method integrating transfer learning and peak load adaptive identification, step S1 specifically includes the following process:

[0012] S11, historical load data preprocessing: preprocess the load data, including missing value filling, outlier removal and normalization;

[0013] S12, Transfer Learning Loss Function Construction: A transfer learning approach is used to take into account the correlation between each busbar and thus improve the prediction accuracy. In the transfer learning framework, the goal is to improve the prediction accuracy of the target busbar by transferring the knowledge of the source busbar to the target busbar. First, the source busbar data is pre-trained to obtain the source busbar prediction model. This prediction model is then applied to the target busbar and the parameters of the target task are fine-tuned to adapt to the target busbar.

[0014] S13, weight update: After each training iteration, the weights are updated based on the error of the current model. As training progresses, the source bus prediction model gradually learns the contribution of each source bus to the target bus, thereby dynamically adjusting the weights.

[0015] S14, model training: By optimizing the transfer learning loss function, the prediction model is trained to improve the prediction effect of the target bus.

[0016] In the above-mentioned bus peak load prediction method integrating transfer learning and peak load adaptive identification, in step S11, the load data L(t) is defined as:

[0017] L(t)=f(historical_load(t),data_features(t),weather_features(t)) (21)

[0018] In formula (1), f represents the data processing function, historical_load(t) is the historical load data, data_features(t) and weather_features(t) are date features and meteorological features, respectively.

[0019] In the above-mentioned bus peak load forecasting method integrating transfer learning and peak load adaptive identification, in step S12, the specific process of constructing the transfer learning loss function is as follows:

[0020] S121, assuming that the influence of each source bus i on the target bus is Then the total loss function can be expressed as:

[0021]

[0022] In formula (2), is the time series of the ith source bus, is the actual load of the i-th source bus, is the predicted value of the busbar;

[0023] S122, loss function L of the target task target Defined according to the standard mean square error loss function:

[0024]

[0025] In formula (3), T target is the time series of the target bus, L target (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus;

[0026] S123, the final combination of transfer learning loss functions, not only includes the loss L of the target task target , which also contains the weighted sum of the losses of all source buses:

[0027]

[0028] In formula (4), ω i is the weight, and λ is a hyperparameter that controls the balance between the source bus and target bus tasks.

[0029] In the above-mentioned bus peak load prediction method integrating transfer learning and peak load adaptive identification, in step S13, the weight update process is implemented by the gradient descent method, and the update rule is:

[0030]

[0031] In formula (5), η is the learning rate, L is the total loss function, is the loss function over weights The updated weights reflect the contribution of each source bus to the training of the target bus.

[0032] In the above-mentioned bus peak load prediction method integrating transfer learning and peak load adaptive identification, in step S14, the specific process of model training is as follows:

[0033] S141, initialization model: use the data of each source bus and target bus to train the initial source bus and target bus prediction model:

[0034]

[0035] In formula (6) and formula (7), represents the parameter optimization result of the i-th source bus, θ represents the various parameters of the i-th source bus, represents the loss function of the i-th source bus under the parameter θ; represents the parameter optimization result of the target busbar, θ T0 Indicates the various parameters of the target bus, L target (θ T0 ) indicates that the target bus is at parameter θ T0 The loss function under

[0036] S142, fine-tuning the target bus: using the transfer learning loss function L transfer Fine-tune the target bus and optimize the top-level parameters of the target bus:

[0037]

[0038] In formula (8), ω i is the weight, λ is the hyperparameter that controls the balance between the source bus and target bus tasks, Indicates that the i-th source bus is in parameter The loss function under target (θ T ) indicates that the target bus is at parameter θ T The loss function under Indicates the parameter fine-tuning result of the target busbar, θ T Indicates various parameters of the target bus.

[0039] In the above-mentioned bus peak load forecasting method integrating transfer learning and peak load adaptive identification, step S2 specifically includes the following process:

[0040] S21, Volatility Measurement and Dynamic Adjustment Strategy: This paper introduces a dynamic peak load period identification technology based on an adaptive threshold method, which uses dynamic thresholds to identify peak load periods.

[0041] S22, peak load period and regular load period identification: Use dynamic thresholds to dynamically identify peak load periods and regular load periods in load fluctuations, and maintain high adaptability under different load fluctuation environments.

[0042] In the above-mentioned bus peak load prediction method integrating transfer learning and peak load adaptive identification, in step S21, the calculation formula of the dynamic threshold θ(t) is:

[0043] θ(t)=μ(t)+k(t)·σ(t) (29)

[0044] In formula (9), μ(t) is the moving average of the load data, which represents the long-term trend of the load; σ(t) is the standard deviation of the load data, which reflects the load volatility; k is a dynamic adjustment coefficient used to balance the impact of volatility on the threshold. In order to ensure efficient tracking of load fluctuations, the following method is used to calculate the standard deviation σ(t) and moving average μ(t) within the time window t:

[0045] The standard deviation σ(t) is calculated as:

[0046]

[0047] In formula (10), L(i) is the load value at the i-th moment, and N is the length of the defined time window;

[0048] The moving average μ(t) is calculated as follows:

[0049]

[0050] In formula (11), L(i) is the load value at the i-th moment, and N is the length of the defined time window;

[0051] Through the two statistics of standard deviation σ(t) and moving average μ(t), the dynamic threshold θ(t) adapts to the different fluctuations in load data and accurately identifies peak load periods and normal load periods;

[0052] Dynamic threshold update strategy: The volatility of load data may change with seasonal changes, meteorological factors, or load characteristics. k is dynamically adjusted through the following strategy:

[0053]

[0054] In formula (12), η is a constant coefficient used to control the sensitivity of dynamic adjustment; PeakRate(t) is the proportion of peak load periods in the load data, indicating the ratio of peak load to total load; StdDev(t) is the standard deviation of the load data; γ is a weight coefficient used to adjust the influence of the standard deviation on the dynamic adjustment coefficient k;

[0055] The proportion of peak load periods in load data is defined as:

[0056]

[0057] In formula (13), T peak is the peak load period, T total is the total load period;

[0058] The dynamic adjustment mechanism ensures that when the load fluctuation is large, the threshold value during the peak load period will be more relaxed to capture more fluctuation information; while when the load fluctuation is small, the threshold value will be stricter to avoid misidentification.

[0059] In the above bus peak load prediction method integrating transfer learning and peak load adaptive identification, in step S22, a dynamic threshold θ(t) is used, and the load period division can be completed by the following rules:

[0060] Identification of peak load period: For each moment t, if L(t)>θ(t), then the moment is judged to be the peak load period t peak ,Right now:

[0061] t peak ={t|L(t)>θ(t)} (34)

[0062] Identification of normal load period: For each moment t, if L(t)≤θ(t), then the moment is judged to be normal load period t normal ,Right now:

[0063] t normal ={t|L(t)≤θ(t)} (35)

[0064] In equations (14) and (15), L(t) is the load value at the tth moment, and θ(t) is the dynamic threshold at the tth moment. Through equations (14) and (15), the peak load period and the normal load period in the load fluctuation can be dynamically identified, and a high degree of adaptability can be maintained under different load fluctuation environments.

[0065] In the above-mentioned bus peak load prediction method integrating transfer learning and peak load adaptive identification, in step S3, the specific loss function is designed as follows:

[0066] S31, load loss function design for each period: for the normal load period, the mean square error loss function L is used. normal :

[0067]

[0068] In formula (16), T normal That is, the normal load period t normal The set where L true (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus;

[0069] In order to enhance learning during peak load periods, a weighted mean square error loss function is used, and the weight function β(t) is defined as:

[0070] β(t)=α(Ltrue (t)+ε) (37)

[0071] In formula (17), L true (t) is the actual load of the target bus, and α and ε are coefficients.

[0072] Thus the peak load loss function L peak for:

[0073]

[0074] In formula (18), T peak That is, the peak load period t peak The set where β(t) is the weight function, L true (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus;

[0075] S32, customized loss function for prediction model fine-tuning: the total loss function L combines the normal load loss and peak load loss, and introduces a high-order smoothing term L smooth , to reduce load fluctuations:

[0076] L=a1·L normal +a2·L peak +a3·L smooth (39)

[0077] In formula (19), a1, a2 and a3 are hyperparameters that control the relative importance of each part of the loss; the smoothing term L smooth Use the following formula:

[0078]

[0079] In formula (20), κ is the smoothing coefficient, is the derivative of load forecasting.

[0080] The present invention's bus peak load forecasting method, which integrates transfer learning and peak load adaptive identification, significantly improves the target bus forecast accuracy through multi-source bus transfer learning. The adaptive threshold recognition mechanism enables dynamic determination of peak periods that are sensitive to load fluctuations, enhancing the model's perception of critical load periods. The customized loss function strengthens the peak error weight and, combined with smoothing constraints, effectively suppresses drastic fluctuations in the forecast curve. The synergistic effect of the above technical means ultimately achieves significant improvements in peak load forecast accuracy and stability while improving overall forecast accuracy, meeting the power grid's actual demand for "high precision and strong robustness" in load forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1This is a flow chart of the bus peak load prediction method that integrates transfer learning and peak load adaptive identification of the present invention. DETAILED DESCRIPTION

[0082] In order to enable those skilled in the art to better understand the technical solution of the present invention, the specific implementation methods thereof are described in detail below with reference to the accompanying drawings:

[0083] See also Figure 1 The best embodiment of the present invention is a bus peak load prediction method integrating transfer learning and peak load adaptive identification, comprising the following steps:

[0084] S1, bus net load prediction transfer learning framework: Build a dynamic weight-driven transfer learning framework to achieve accurate transfer of multi-source bus knowledge to the target bus through weight update and model training of the prediction model.

[0085] S11, historical load data preprocessing: preprocess the load data L(t), including filling missing values, removing outliers and normalizing; select features closely related to load forecasting, including historical load data, date features and meteorological features, as inputs to the load forecasting model, and carry out data preprocessing tasks. The load data L(t) is defined as:

[0086] L(t)=f(historical_load(t),data_features(t),weather_features(t)) (41)

[0087] In formula (1), f represents the data processing function, historical_load(t) is the historical load data, data_features(t) and weather_features(t) are date features and meteorological features, respectively.

[0088] S12, construction of transfer learning loss function: Since a region usually contains multiple bus loads, and there is temporal and spatial correlation between the electricity consumption of multiple bus loads, it is considered to use a transfer learning method to take into account the correlation between each bus, thereby improving the prediction accuracy. In the transfer learning framework, the goal is to improve the prediction accuracy of the target bus by transferring the knowledge of the source bus (i.e., other buses) to the target bus. First, pre-training is performed on the source bus data to obtain a prediction model for the source bus, and then the model is applied to the target bus, and the parameters of the target task are fine-tuned to adapt to the target bus. The specific process of constructing the transfer learning loss function is as follows:

[0089] S121, assuming that the influence of each source bus i on the target bus is Then the total loss function can be expressed as:

[0090]

[0091] In formula (2), is the time series of the ith source bus, is the actual load of the i-th source bus, is the predicted value of the busbar;

[0092] S122, similarly, the loss function L of the target task target Defined according to the standard mean square error (MSE) loss function:

[0093]

[0094] In formula (3), T target is the time series of the target bus, L target (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus;

[0095] S123, the final combination of transfer learning loss functions, not only includes the loss L of the target task target , which also contains the weighted sum of the losses of all source buses:

[0096]

[0097] In formula (4), ω i is the weight, and λ is a hyperparameter that controls the balance between the source bus and target bus tasks.

[0098] S13, weight update: After each training iteration, the weights are updated according to the error of the current model. As the training progresses, the source bus prediction model gradually learns the contribution of each source bus to the target bus, thereby dynamically adjusting the weights. The weight update process is implemented by the gradient descent method, and the update rule is:

[0099]

[0100] In formula (5), η is the learning rate, L is the total loss function, is the loss function over weights The updated weights reflect the contribution of each source bus to the training of the target bus.

[0101] S14, model training: by optimizing the transfer learning loss function L transfer , train the prediction model to improve the prediction effect of the target bus. The specific process of model training is as follows:

[0102] S141, initialization model: use the data of each source bus and target bus to train the initial source bus and target bus prediction model:

[0103]

[0104] In formula (6) and formula (7), represents the parameter optimization result of the i-th source bus, θ represents the various parameters of the i-th source bus, Represents the loss function of the i-th source bus under the parameter θ. represents the parameter optimization result of the target busbar, θ T0 Indicates the various parameters of the target bus, L target (θ T0 ) indicates that the target bus is at parameter θ T0 The loss function below.

[0105] S142, fine-tuning the target bus: using the transfer learning loss function L transfer Fine-tune the target bus and optimize the top-level parameters of the target bus:

[0106]

[0107] In formula (8), ω i is the weight, λ is the hyperparameter that controls the balance between the source bus and target bus tasks, Indicates that the i-th source bus is in parameter The loss function under target (θ T ) indicates that the target bus is at parameter θ T The loss function under Indicates the parameter fine-tuning result of the target busbar, θ T Indicates various parameters of the target bus.

[0108] S2, dynamic peak load identification based on adaptive threshold method: Introducing an adaptive threshold mechanism based on load volatility and peak ratio to dynamically identify peak and regular load periods. The specific process includes the following:

[0109] S21, Volatility measurement and dynamic adjustment strategy: The existing forecasting methods have limited accuracy in predicting instantaneous peak loads. To improve the accuracy of peak load identification in load forecasting, this step introduces a dynamic peak load period identification technology based on an adaptive threshold method. The peak load period is identified by using a dynamic threshold θ(t). The calculation formula of the dynamic threshold θ(t) is:

[0110] θ(t)=μ(t)+k(t)·σ(t) (49)

[0111] In formula (9), μ(t) is the moving average of the load data, which represents the long-term trend of the load; σ(t) is the standard deviation of the load data, which reflects the load volatility; k is a dynamic adjustment coefficient used to balance the impact of volatility on the threshold. In order to ensure efficient tracking of load fluctuations, the following method is used to calculate the standard deviation σ(t) and moving average μ(t) within the time window t:

[0112] (1) The formula for calculating the standard deviation σ(t) is:

[0113]

[0114] In formula (10), L(i) is the load value at the i-th moment, and N is the length of the defined time window;

[0115] (2) The calculation formula of moving average μ(t) is:

[0116]

[0117] In formula (11), L(i) is the load value at the i-th moment, and N is the length of the defined time window;

[0118] Through the two statistics of standard deviation σ(t) and moving average μ(t), the dynamic threshold θ(t) adapts to the different fluctuations in load data and accurately identifies peak load periods and normal load periods;

[0119] (3) Dynamic threshold update strategy: The volatility of load data will change with seasonal changes, meteorological factors or load characteristics. k is dynamically adjusted through the following strategy:

[0120]

[0121] In formula (12), η is a constant coefficient used to control the sensitivity of dynamic adjustment; PeakRate(t) is the proportion of peak load periods in the load data, indicating the ratio of peak load to total load; StdDev(t) is the standard deviation of the load data; γ is a weight coefficient used to adjust the influence of the standard deviation on the dynamic adjustment coefficient k;

[0122] The proportion of peak load periods in load data is defined as:

[0123]

[0124] In formula (13), T peak is the peak load period, T total is the total load period;

[0125] This dynamic adjustment mechanism ensures that when the load fluctuation is large, the threshold value during the peak load period will be more relaxed to capture more fluctuation information; while when the load fluctuation is small, the threshold value will be stricter to avoid misidentification.

[0126] S22, Peak Load Period and Normal Load Period Identification: Use the dynamic threshold θ(t) to dynamically identify the peak load period and normal load period in the load fluctuation, and maintain a high degree of adaptability under different load fluctuation environments. Using the dynamic threshold θ(t), the load period division can be completed according to the following rules:

[0127] Identification of peak load period: For each moment t, if L(t)>θ(t), then the moment is judged to be the peak load period t peak ,Right now:

[0128] t peak ={t|L(t)>θ(t)} (54)

[0129] Identification of normal load period: For each moment t, if L(t)≤θ(t), then the moment is judged to be normal load period t normal ,Right now:

[0130] t normal ={t|L(t)≤θ(t)} (55)

[0131] In formulas (14) and (15), L(t) is the load value at the tth moment, and θ(t) is the dynamic threshold at the tth moment. Through formulas (14) and (15), the peak load period and the normal load period in the load fluctuation can be dynamically identified, and a high degree of adaptability can be maintained under different load fluctuation environments.

[0132] S3, Customized Loss Function and Model Fine-tuning: A customized loss function that emphasizes peak error weighting is designed and used for model fine-tuning to improve bus load forecasting accuracy during critical periods. After determining peak and regular load periods, a customized loss function is introduced to further improve forecast accuracy during peak load periods, with a focus on emphasizing error weighting during peak load periods.

[0133] The specific loss function is designed as follows:

[0134] S31, load loss function design for each period: for the normal load period, the mean square error loss function L is used. normal :

[0135]

[0136] In formula (16), T normal That is, the normal load period t normal The set where Ltrue (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus;

[0137] In order to enhance learning during peak load periods, the weighted mean square error (W-MSE) loss function is adopted, and the weight function β(t) is defined as:

[0138] β(t)=α(L true (t)+ε) (57)

[0139] In formula (17), L true (t) is the actual load of the target bus, and α and ε are coefficients.

[0140] Thus the peak load loss function L peak for:

[0141]

[0142] In formula (18), T peak That is, the peak load period t peak The set where β(t) is the weight function, L true (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus;

[0143] S32, customized loss function for prediction model fine-tuning: the total loss function L combines the normal load loss and peak load loss, and introduces a high-order smoothing term L smooth , to reduce load fluctuations:

[0144] L=a1·L normal +a2·L peak +a3·L smooth (59)

[0145] In formula (19), a1, a2 and a3 are hyperparameters that control the relative importance of each part of the loss; the smoothing term L smooth Use the following formula:

[0146]

[0147] In formula (20), κ is the smoothing coefficient, is the derivative of load forecasting.

[0148] Furthermore, the customized loss function is used to fine-tune the parameters of the prediction model in step S1, thereby improving the prediction accuracy during peak load periods and providing valuable guidance for grid supply security.

[0149] In summary, the bus peak load prediction method of the present invention, which integrates transfer learning and peak load adaptive identification, innovatively constructs a dynamic weight-driven transfer learning framework, realizes the precise transfer of multi-source bus knowledge to the target bus; introduces an adaptive threshold mechanism based on load volatility and peak ratio to dynamically identify peak and regular load periods; further designs a customized loss function that strengthens the peak error weight and uses it for model fine-tuning, thereby significantly improving the load prediction accuracy of the bus in critical periods. The target bus prediction accuracy is significantly improved through multi-source bus transfer learning; the adaptive threshold identification mechanism realizes the dynamic judgment of peak periods that are sensitive to load fluctuations, and enhances the model's perception of critical load periods; the customized loss function strengthens the peak error weight, and combines with smoothing constraints to effectively suppress the violent fluctuations of the prediction curve. The above technical means work synergistically, and ultimately achieve the goal of significantly improving the prediction accuracy and stability of peak load while improving the overall prediction accuracy, meeting the actual needs of the power grid for "high precision and strong robustness" in load forecasting.

[0150] Those skilled in the art should recognize that the above embodiments are merely intended to illustrate the present invention and are not intended to limit the present invention. As long as they are within the spirit of the present invention, any changes or modifications to the above embodiments will fall within the scope of the claims of the present invention.

Claims

1. A bus peak load prediction method integrating transfer learning and peak load adaptive identification, characterized in that: The following steps are involved: S1: Build a dynamic weight-driven transfer learning framework to accurately transfer multi-source bus knowledge to the target bus through weight updates and model training of the prediction model; S2, introduces an adaptive threshold mechanism based on load volatility and peak proportion to dynamically identify peak and regular load periods; S3, designs a customized loss function that strengthens the peak error weight and uses it for model fine-tuning to improve the load forecasting accuracy of the bus during critical periods.

2. A bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 1, characterized in that: Step S1 specifically includes the following process: S11, historical load data preprocessing: preprocess the load data, including missing value filling, outlier removal and normalization; S12, Transfer Learning Loss Function Construction: A transfer learning approach is used to take into account the correlation between each busbar and thus improve the prediction accuracy. In the transfer learning framework, the goal is to improve the prediction accuracy of the target busbar by transferring the knowledge of the source busbar to the target busbar. First, the source busbar data is pre-trained to obtain the source busbar prediction model. This prediction model is then applied to the target busbar and the parameters of the target task are fine-tuned to adapt to the target busbar. S13, weight update: After each training iteration, the weights are updated based on the error of the current model. As training progresses, the source bus prediction model gradually learns the contribution of each source bus to the target bus, thereby dynamically adjusting the weights. S14, model training: By optimizing the transfer learning loss function, the prediction model is trained to improve the prediction effect of the target bus.

3. The bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 2 is characterized in that: In step S11, the load data L(t) is defined as: L(t)=f(historical_load(t),data_features(t),weather_features(t)) (1) In formula (1), f represents the data processing function, historical_load(t) is the historical load data, data_features(t) and weather_features(t) are date features and meteorological features, respectively.

4. A bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 3, characterized in that: In step S12, the specific process of constructing the transfer learning loss function is as follows: S121, assuming that the influence of each source bus i on the target bus is Then the total loss function can be expressed as: In formula (2), is the time series of the ith source bus, is the actual load of the i-th source bus, is the predicted value of the busbar; S122, loss function L of the target task target Defined according to the standard mean square error loss function: In formula (3), T target is the time series of the target bus, L target (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus; S123, the final combination of transfer learning loss functions, not only includes the loss L of the target task target , which also contains the weighted sum of the losses of all source buses: In formula (4), ω i is the weight, and λ is a hyperparameter that controls the balance between the source bus and target bus tasks.

5. The bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 4 is characterized in that: In step S13, the weight update process is implemented by the gradient descent method, and the update rule is: In formula (5), η is the learning rate, L is the total loss function, is the loss function over weights The updated weights reflect the contribution of each source bus to the training of the target bus.

6. A bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 5, characterized in that: In step S14, the specific process of model training is as follows: S141, initialization model: use the data of each source bus and target bus to train the initial source bus and target bus prediction model: In formula (6) and formula (7), represents the parameter optimization result of the i-th source bus, θ represents the various parameters of the i-th source bus, represents the loss function of the i-th source bus under the parameter θ; represents the parameter optimization result of the target busbar, θ T0 Indicates the various parameters of the target bus, L target (θ T0 ) indicates that the target bus is at parameter θ T0 The loss function under S142, fine-tuning the target bus: using the transfer learning loss function L transfer Fine-tune the target bus and optimize the top-level parameters of the target bus: In formula (8), ω i is the weight, λ is the hyperparameter that controls the balance between the source bus and target bus tasks, Indicates that the i-th source bus is in parameter The loss function under target (θ T ) indicates that the target bus is at parameter θ T The loss function under Indicates the parameter fine-tuning result of the target busbar, θ T Indicates various parameters of the target bus.

7. The bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 1 is characterized in that: Step S2 specifically includes the following process: S21, Volatility Measurement and Dynamic Adjustment Strategy: This paper introduces a dynamic peak load period identification technology based on an adaptive threshold method, which uses dynamic thresholds to identify peak load periods. S22, peak load period and regular load period identification: Use dynamic thresholds to dynamically identify peak load periods and regular load periods in load fluctuations, and maintain high adaptability under different load fluctuation environments.

8. The bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 7 is characterized in that: In step S21, the calculation formula of the dynamic threshold θ(t) is: θ(t)=μ(t)+k(t)·σ(t) (9) In formula (9), μ(t) is the moving average of the load data, which represents the long-term trend of the load; σ(t) is the standard deviation of the load data, which reflects the load volatility; k is a dynamic adjustment coefficient used to balance the impact of volatility on the threshold. In order to ensure efficient tracking of load fluctuations, the following method is used to calculate the standard deviation σ(t) and moving average μ(t) within the time window t: The standard deviation σ(t) is calculated as: In formula (10), L(i) is the load value at the i-th moment, and N is the length of the defined time window; The moving average μ(t) is calculated as follows: In formula (11), L(i) is the load value at the i-th moment, and N is the length of the defined time window; Through the two statistics of standard deviation σ(t) and moving average μ(t), the dynamic threshold θ(t) adapts to the different fluctuations in load data and accurately identifies peak load periods and normal load periods; Dynamic threshold update strategy: The volatility of load data may change with seasonal changes, meteorological factors, or load characteristics. k is dynamically adjusted through the following strategy: In formula (12), η is a constant coefficient used to control the sensitivity of dynamic adjustment; PeakRate(t) is the proportion of peak load periods in the load data, indicating the ratio of peak load to total load; StdDev(t) is the standard deviation of the load data; γ is a weight coefficient used to adjust the influence of the standard deviation on the dynamic adjustment coefficient k; The proportion of peak load periods in load data is defined as: In formula (13), T peak is the peak load period, T total is the total load period; The dynamic adjustment mechanism ensures that when the load fluctuation is large, the threshold value during the peak load period will be more relaxed to capture more fluctuation information; while when the load fluctuation is small, the threshold value will be stricter to avoid misidentification.

9. The bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 8 is characterized in that: In step S22, using the dynamic threshold θ(t), the load period division can be completed according to the following rules: Identification of peak load period: For each moment t, if L(t)>θ(t), then the moment is judged to be the peak load period t peak ,Right now: t peak ={t|L(t)>θ(t)} (14) Identification of normal load period: For each moment t, if L(t)≤θ(t), then the moment is judged to be normal load period t normal ,Right now: t normal ={t|L(t)≤θ(t)} (15) In Equations (14) and (15), L(t) is the load value at the tth moment, and θ(t) is the dynamic threshold at the tth moment. Through Equations (14) and (15), the peak load period and the normal load period in the load fluctuation can be dynamically identified, and a high degree of adaptability can be maintained under different load fluctuation environments.

10. The bus peak load prediction method integrating transfer learning and peak load adaptive identification according to claim 1 is characterized in that: In step S3, the specific loss function is designed as follows: S31, load loss function design for each period: for the normal load period, the mean square error loss function L is used. normal : In formula (16), T normal That is, the normal load period t normal The set where L true (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus; In order to enhance learning during peak load periods, a weighted mean square error loss function is used, and the weight function β(t) is defined as: β(t)=α(L true (t)+ε) (17) In formula (17), L true (t) is the actual load of the target bus, α and ε are coefficients; Thus the peak load loss function L peak for: In formula (18), T peak That is, the peak load period t peak The set where β(t) is the weight function, L true (t) is the actual load of the target bus, L pred (t) is the predicted value of the target bus; S32, customized loss function for prediction model fine-tuning: the total loss function L combines the normal load loss and peak load loss, and introduces a high-order smoothing term L smooth , to reduce load fluctuations: <h2 style=";text-align:left;direction:ltr">L = a1 L<h2 style=";text-align:left;direction:ltr"> normal <h2 style=";text-align:left;direction:ltr"> +a2·L<h2 style=";text-align:left;direction:ltr"> peak <h2 style=";text-align:left;direction:ltr"> +a3·L<h2 style=";text-align:left;direction:ltr"> smooth <h2 style=";text-align:left;direction:ltr"> (19) In formula (19), a1, a2 and a3 are hyperparameters that control the relative importance of each part of the loss; the smoothing term L smooth Use the following formula: In formula (20), κ is the smoothing coefficient, is the derivative of load forecasting.

Images