Baseline load accurate generation method and system of source network load storage integrated park

Abstract

The invention discloses a baseline load accurate generation method and system for a source network load storage integrated park, and the method comprises the steps: carrying out the pre-separation processing of a gateway power curve according to a directly controllable load, a non-directly controllable load and a distributed power supply, and building a data set according to a non-directly controllable load curve and an exogenous feature; and performing data reconstruction on the data set according to time to obtain a reconstructed data set, training the improved TGDA model by using the reconstructed data set to obtain a predicted non-directly-controllable load curve, and finally generating a complete baseline load curve. According to the method, accurate prediction of the park baseline load is realized through effective screening of external source features, reasonable reconstruction of a data structure and collaborative modeling of the hybrid neural network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of park baseline load, in particular to a baseline load accurate generation method and system for a source-grid-load-storage integrated park. BACKGROUND

[0002] With the global energy structure accelerating towards low-carbonization and intelligentization, the flexible regulation and control capability of demand-side resources of the power system puts forward higher requirements. Demand response (DR) as a key technical means to realize the dynamic balance of power supply and demand and improve the level of renewable energy consumption, its implementation effect depends to a large extent on the generation capability of high-precision baseline load. Baseline load is defined as the expected power consumption curve of users without taking any demand response measures, and its accuracy is directly related to the economic settlement mechanism of DR project, user participation enthusiasm and the formulation of power dispatching strategy.

[0003] With the continuous development of new power systems, the penetration rate of distributed new energy and energy storage devices in the power grid continues to rise, and more and more parks are gradually evolving into integrated parks that integrate power sources, power grids, loads and energy storage devices, i.e. source-grid-load-storage integrated parks. Such parks usually have complex load composition, containing a large number of mixed loads, and the adjustable resources are diverse in types and different in adjustment methods, resulting in significant noise in the overall load data and large differences in user power consumption behavior, thereby significantly increasing the technical difficulty of baseline load accurate modeling.

[0004] Traditional baseline load generation methods are mainly based on statistical models such as rolling mean method and median method, although the calculation process is simple and efficient, but when dealing with source-grid-load-storage integrated parks with high complexity, there are often large errors, and the actual prediction deviation often exceeds 20%, which is difficult to meet the basic requirements of accuracy for engineering applications. In addition, some existing researches attempt to use machine learning methods (such as LSTM, XGBoost, etc.) for modeling, although these methods can capture the nonlinear fluctuation characteristics of the load to some extent, but due to their relatively simple network architecture and limited feature extraction capability, they still show obvious limitations when dealing with complex load data of source-grid-load-storage integrated parks.

[0005] On the other hand, although load decomposition technology (such as non-intrusive load monitoring NILM) has the ability to separate adjustable load and baseline load, its implementation relies on high-precision metering equipment, which brings high hardware cost and limits its popularization and application in large-scale scenarios.

[0006] Furthermore, traditional baseline load generation methods often rely on direct forecasting of load at the critical point. However, integrated energy-grid-load-storage parks typically contain a large number of directly controllable loads (such as energy storage systems and controllable charging piles). The actual operating data of these loads can be directly obtained through the local management system without additional forecasting. Modeling these loads together with uncontrollable loads not only introduces unnecessary random interference but also significantly increases the difficulty of model prediction and reduces overall prediction accuracy. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to provide a method and system for accurately generating the baseline load of an integrated power generation, grid, load and storage park, so as to achieve accurate prediction of the park's baseline load.

[0008] Technical solution: The method for accurately generating baseline load in an integrated source-grid-load-storage park as described in this invention includes the following steps:

[0009] Obtain the gate power curve of the park, and perform pre-separation processing on the gate power curve according to the directly controllable load, the non-directly controllable load and the distributed power source to obtain the non-directly controllable load curve and the distributed power source output curve.

[0010] A dataset was established based on the non-controllable load curve and external characteristics, and the dataset was reconstructed according to time to obtain the reconstructed dataset; the external characteristics were obtained by screening relevant characteristic variables using the mutual information method.

[0011] An improved TGDA model was trained using a reconstructed dataset, and the improved TGDA model was used to obtain the predicted non-controllable load curve; the feature mining layer of the improved TGDA model embeds an attention mechanism module after the TCN network structure.

[0012] The predicted non-controllable load curve, the distributed power generation output curve, and the expected controllable load curve are superimposed to obtain a complete baseline load curve.

[0013] Furthermore, the data reconstruction of the dataset according to time includes:

[0014] The historical load data of several days before the generation date of the power curve at the gateway are concatenated with the corresponding external features on the generation date in chronological order and flattened into a one-dimensional feature vector, which serves as one dimension of the reconstructed dataset.

[0015] Based on the one-dimensional feature vector, a sliding window with a step size of one day is used to continuously divide the data to obtain another dimension of the reconstructed dataset.

[0016] Furthermore, the improved TGDA model also includes: embedding an attention mechanism module between the two GRU layers in the temporal correlation establishment layer of the improved TGDA model.

[0017] Furthermore, the complete baseline load curve P 基线 The calculation method is as follows:

[0018] P 基线 =P 非可直控_预测 -P 分布式电源 +P 可直控_原 ;

[0019] Among them, P 非可直控_预测 To predict non-controllable load curves; P 分布式电源 The output curve of the distributed power source obtained from the pre-separation process; P 可直控_原 This represents the expected controllable load curve when no active regulation is implemented.

[0020] Furthermore, the directly controllable load includes load devices whose power output can be freely adjusted at any time period, and the non-directly controllable load includes non-adjustable fixed loads and load devices that can be adjusted to a limited extent.

[0021] The baseline load precise generation system for the integrated source-grid-load-storage park described in this invention includes:

[0022] The power curve pre-separation unit is used to obtain the gate power curve of the park and pre-separate the gate power curve according to the directly controllable load, the non-directly controllable load and the distributed power source to obtain the non-directly controllable load curve and the distributed power source output curve.

[0023] The dataset creation unit is used to create a dataset based on the non-controllable load curve and external characteristics, and to reconstruct the dataset according to time to obtain a reconstructed dataset; the external characteristics are obtained by screening relevant feature variables using the mutual information method;

[0024] The TGDA modeling unit is used to train and improve the TGDA model using the reconstructed dataset, and to obtain the predicted non-controllable load curve using the improved TGDA model; the feature mining layer of the improved TGDA model embeds an attention mechanism module after the TCN network structure.

[0025] The baseline load curve generation unit is used to overlay the predicted non-controllable load curve, the distributed power output curve, and the expected controllable load curve to obtain a complete baseline load curve.

[0026] Furthermore, in the dataset creation unit, the data reconstruction of the dataset according to time includes:

[0027] The historical load data of several days before the generation date of the power curve at the gateway are concatenated with the corresponding external features on the generation date in chronological order and flattened into a one-dimensional feature vector, which serves as one dimension of the reconstructed dataset.

[0028] Based on the one-dimensional feature vector, a sliding window with a step size of one day is used to continuously divide the data to obtain another dimension of the reconstructed dataset.

[0029] Furthermore, the improved TGDA model modeling unit also includes: embedding an attention mechanism module between the two GRU layers in the temporal correlation establishment layer of the improved TGDA model.

[0030] Furthermore, in the baseline load curve generation unit, the complete baseline load curve P 基线 The calculation method is as follows:

[0031] P 基线 =P 非可直控_预测 -P 分布式电源 +P 可直控_原 ;

[0032] Among them, P 非可直控_预测 To predict non-controllable load curves; P 分布式电源 The output curve of the distributed power source obtained from the pre-separation process; P 可直控_原 This represents the expected controllable load curve when no active regulation is implemented.

[0033] Furthermore, in the power curve pre-separation unit, the directly controllable load includes load devices whose power output can be freely adjusted at any time period, and the non-directly controllable load includes non-adjustable fixed loads and load devices that can be adjusted to a limited extent.

[0034] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the computer program is loaded onto the processor, it implements the baseline load accurate generation method for the integrated source-grid-load-storage park.

[0035] The computer program product of the present invention includes a computer program that, when executed by a processor, implements a method for accurately generating the baseline load of the integrated source-grid-load-storage park.

[0036] Beneficial effects: Compared with existing technologies, the advantages of this invention are: through effective screening of external features, reasonable reconstruction of data structures, and collaborative modeling of hybrid neural networks, accurate prediction of the baseline load of the park is achieved. This method not only improves prediction accuracy but also takes into account engineering practicality and deployment feasibility, providing reliable technical support for load management in integrated power generation, grid, load, and storage parks. Attached Figure Description

[0037] Figure 1 This is a flowchart of the baseline load accurate generation method of the present invention.

[0038] Figure 2This is a flowchart of the baseline load accurate generation method according to an embodiment of the present invention.

[0039] Figure 3 This is a comparison chart of baseline load prediction results under a single-shift system according to an embodiment of the present invention.

[0040] Figure 4 This is a comparison chart of baseline load prediction results under a two-shift system according to an embodiment of the present invention.

[0041] Figure 5 This is a comparison chart of baseline load prediction results under a single-shift system according to an embodiment of the present invention.

[0042] Figure 6 This is a comparison chart of baseline load prediction results under a three-shift system according to an embodiment of the present invention.

[0043] Figure 7 This is a comparison chart of baseline load prediction results for a certain winter day according to an embodiment of the present invention.

[0044] Figure 8 This is a comparison chart of baseline load prediction results for a certain day in spring according to an embodiment of the present invention.

[0045] Figure 9 This is a comparison chart of baseline load prediction results for a certain day in summer according to an embodiment of the present invention.

[0046] Figure 10 This is a comparison chart of baseline load prediction results for a certain autumn day according to an embodiment of the present invention.

[0047] Figure 11 This is a comparison chart of baseline load prediction results for a sunny day according to an embodiment of the present invention.

[0048] Figure 12 This is a comparison chart of baseline load prediction results for a rainy day according to an embodiment of the present invention. Detailed Implementation

[0049] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0050] like Figure 1 As shown, the method for accurately generating the baseline load of the integrated power generation, grid, load, and storage park includes four steps: pre-separation of the power curve at the park's access point, construction and partitioning of the baseline load dataset, TGDA model training and inference, and final generation of the baseline load curve. This method not only improves prediction accuracy but also takes into account engineering practicality and deployment feasibility, providing reliable technical support for load management in integrated power generation, grid, load, and storage parks. The following section will combine... Figure 2 The method described in this invention will be described in detail.

[0051] S1. Given the high complexity of the threshold power curve, a pre-separation process is adopted for the threshold power curve in the park to obtain the non-directly controllable load curve.

[0052] Specifically, this invention pre-separates the separable portion from the threshold power curve, thereby effectively reducing the noise level and volatility of the overall curve and improving the accuracy and stability of load forecasting. Considering the load composition characteristics of integrated power generation, grid, load, and storage parks, this invention classifies them into three categories by definition: directly controllable loads, non-directly controllable loads, and distributed power sources.

[0053] Directly controllable loads refer to load devices whose power output can be freely adjusted at any time, such as energy storage systems and some charging piles with intelligent control functions. Because such loads are fully controllable, their daily output curves can be directly obtained through relevant management equipment, exhibiting high real-time performance and accuracy.

[0054] Non-directly controllable loads include non-adjustable fixed loads and some load equipment that can only be adjusted to a limited extent, such as air conditioning systems and cold storage equipment that only support start-stop control; charging piles that are affected by user behavior and can only be used through electricity price guidance; or industrial loads that need to have their operating time set in advance according to production plans. Since these loads cannot be directly separated by existing equipment, in order to avoid increasing additional hardware costs, this invention uses the proposed TGDA deep learning model to model and predict them.

[0055] Distributed power sources mainly refer to energy devices such as distributed photovoltaic power generation systems configured within the industrial park. Although their power generation is not considered part of the load, it has a direct impact on the threshold power. Therefore, during model training, this portion of power needs to be pre-extracted from the original threshold curve to improve prediction accuracy and model robustness.

[0056] In summary, the purpose of pre-separation of the park's control curve is to obtain the non-directly controllable load curve, and the formula is shown below:

[0057] P 非可直控 =P 关口 -P 可直控 +P 分布式电源 (1)

[0058] Where: P 非可直控 For the required calculation of the non-directly controllable load power, P 关口 The power measured at the gate of the integrated power generation, grid, load, and storage park, P 可直控 For directly controllable load power, P 分布式电源 This refers to the power generation capacity of distributed generation stations.

[0059] S2, based on non-controllable load curve data, constructs and divides a baseline load dataset to obtain a training set and a test set.

[0060] It is specifically divided into three parts.

[0061] (1) Screening of external influencing factors using mutual information method:

[0062] After obtaining the non-controllable load curve data, the mutual information method is used to screen potentially relevant external reference variables, retaining external features that have a significant information contribution to the load curve, and eliminating feature variables with low information contribution or irrelevant features.

[0063] Mutual information is a feature selection method that can simultaneously measure both linear and nonlinear correlations between features. Its core principle is to assess the strength of information interaction by quantifying the degree of interdependence between two random variables. Strictly speaking, for two discrete random variables X and Y, their mutual information is defined as the relative entropy between the product of their joint probability distribution and their marginal probability distributions, i.e.:

[0064]

[0065] Where p(x,y) is the joint probability distribution function of X and Y, and p(x) and p(y) are the marginal probability distribution functions of X and Y, respectively. Mutual information is the intrinsic dependency between the joint distribution of X and Y relative to the joint distribution assuming X and Y are independent. Therefore, mutual information measures dependency as follows: I(X;Y) is true if and only if X and Y are independent random variables. That is, when the two variables are completely independent and have no relationship, the mutual information is 0, while the stronger the correlation between the two variables (whether linear or nonlinear), the closer the mutual information value is to 1.

[0066] In some optional embodiments, after screening external characteristics, the key information retained for the load characteristics of the integrated source-grid-load-storage park includes shift schedule, order density, whether it is a working day, and weather conditions.

[0067] (2) Dataset restructuring:

[0068] After obtaining the amount of external information and historical non-controllable load data, the existing multidimensional variables are reconstructed. The historical load data of the 7 days before the generation date and the corresponding external influence index data of the generation date are concatenated in chronological order and flattened into a one-dimensional feature vector, which serves as one dimension of the reconstructed dataset matrix. Then, while maintaining this data structure, the data is continuously segmented using a sliding window method with a step size of one day to construct another dimension of the dataset, thereby completing the construction of the overall dataset.

[0069] (3) Division of training and test sets:

[0070] After reconstructing the entire dataset, it needs to be divided into training and testing sets to meet the requirements of model training and testing. Typically, the data can be divided into training and testing sets in a 5:1 ratio, meaning 5 / 6 of the total data is used for model training, and the remaining 1 / 6 is used for model validation and parameter tuning, ensuring the model has good generalization ability and practical application performance.

[0071] S3, based on the TGDA model, is trained using the training set and used for inference using the test set.

[0072] Specifically, it is divided into four parts.

[0073] (1) Overall model architecture of TGDA:

[0074] This model is designed based on an overall architecture of "temporal feature mining – temporal feature association construction," employing a TCN network with an attention mechanism to achieve deep extraction of front-end temporal features. Compared to traditional single models (such as LSTM), this method demonstrates stronger capabilities in temporal feature mining. After feature extraction, the obtained temporal features are input into a GRU layer with an attention mechanism. Leveraging the advantages of the GRU network in temporal information modeling, the dynamic modeling capability of sequence data is further enhanced, thereby significantly improving the accuracy and stability of prediction results. The overall model architecture can be found at [reference needed]. Figure 3 .

[0075] (2) TCN feature mining layer with attention mechanism:

[0076] Convolutional Neural Networks (CNNs) have demonstrated remarkable capabilities in image processing and spatial feature extraction. Through their design of local receptive fields and weight sharing, they can efficiently capture local correlations in input data. However, traditional CNNs have certain limitations in processing temporal data. For example, their fixed receptive fields are difficult to flexibly adapt to sequence structures of different lengths, and they lack the ability to effectively model dynamic dependencies in the time dimension.

[0077] To address these issues, Temporal Convolutional Networks (TCNs) were developed. TCNs can be seen as an extension of CNNs for temporal tasks. They ensure the temporal order of the model through causal convolution and gradually expand the receptive field using dilated convolution, thus effectively capturing information over longer time spans. Furthermore, TCNs typically incorporate residual connections, which not only alleviate the vanishing or exploding gradient problems that can occur in deep networks but also improve the model's training stability and expressive power. Therefore, TCNs inherently possess strong advantages in temporal feature extraction due to their structural design.

[0078] Based on the aforementioned characteristics, using TCN as the "head" structure of the feature mining module helps extract richer and deeper temporal feature representations from the original input. However, relying solely on the traditional TCN structure for feature extraction still has some shortcomings. First, the various feature components output by TCN are structurally equal, lacking a mechanism to distinguish which features are more critical for subsequent tasks. This "one-size-fits-all" feature representation may hinder the effective utilization of feature information by subsequent temporal modeling layers. Second, because TCN itself has strong feature mining capabilities, it may extract some "pseudo-features" that are irrelevant to the task objective or even misleading in some cases. This redundant or false information may interfere with the learning process of subsequent models and affect overall performance.

[0079] To address the above issues, this invention proposes an improved feature mining architecture—embedding an attention mechanism into the TCN network structure to construct a novel feature mining network with feature filtering capabilities. The core of this method lies in utilizing the weight allocation capability of the attention mechanism to weight and label the numerous feature components generated after the TCN convolutional operations. Specifically, after the TCN completes initial feature extraction, the attention module automatically calculates the importance weight of each feature channel or time step according to the needs of the current task, thereby highlighting key features that positively contribute to the task, suppressing secondary features, and filtering out potentially illusory features.

[0080] In this way, the feature mining network not only retains the powerful temporal modeling capabilities of the original TCN, but also achieves refined filtering and enhancement at the feature representation level, making the output features more representative and interpretable. This optimized feature representation is more conducive to the learning and convergence of subsequent temporal correlation modeling layers, thereby improving the overall system performance.

[0081] (3) GRU temporal association establishment layer with attention mechanism:

[0082] After completing the temporal feature mining, the extracted temporal features are passed to the next key processing stage—the temporal association establishment layer. The main function of this layer is to perform in-depth modeling and association of the extracted temporal features, thereby capturing more complex sequence dependencies. The temporal association establishment layer used in this invention is based on a GRU (Gated Recurrent Unit) network structure and further introduces an attention mechanism to improve the model's efficiency in utilizing hidden layer information.

[0083] GRU is an improved recurrent neural network (RNN) that effectively alleviates the gradient vanishing problem in traditional RNNs and performs well in processing medium-length sequence data. However, when faced with longer input sequences, the single-structure GRU network still has certain limitations, especially when processing information from the beginning of the sequence, where it is prone to "forgetting," meaning the model struggles to effectively retain and utilize information from earlier time points, thus limiting its ability to model global temporal correlations.

[0084] To address this issue, this invention proposes an improved solution: introducing an attention mechanism module between two GRU layers. The core idea of ​​the attention mechanism is to allow the model to selectively focus on more critical time steps in the previous sequence when processing information at the current moment, thereby enhancing the memory and utilization of important historical information. Specifically, the first GRU layer is responsible for initially extracting temporal features from the sequence for preliminary modeling. Subsequently, the attention mechanism weights these features, highlighting those time points that have a greater impact on the current state. The weighted features are then input into the second GRU layer for further temporal modeling, thereby achieving more accurate global information association.

[0085] This two-layer GRU structure, which incorporates an attention mechanism, not only enhances the model's ability to process long sequence data but also strengthens its ability to capture complex temporal patterns, making the entire system more expressive and robust in time series modeling tasks.

[0086] (4) Model input and output:

[0087] Given that baseline loads are typically generated daily with a time granularity generally set to 15 minutes, the deep learning model proposed in this invention defines a single input data set as: historical load data from the 7 days prior to the generation date and exogenous influence index data corresponding to the generation date. These two types of data are concatenated chronologically and flattened into a one-dimensional feature vector as model input. For constructing the overall input dataset, while maintaining this data structure, a sliding window method with a daily step size is used to continuously segment the data. The model output is the non-controllable load curve for the target generation date, totaling 96 time points, corresponding to a sample value every 15 minutes.

[0088] S4 is the baseline load curve calculated and generated after obtaining the predicted curve of the non-controllable load.

[0089] After obtaining the predicted curves for non-directly controllable loads, the previously separated output curves of distributed renewable energy power plants are superimposed and restored according to the original data. For directly controllable loads, the curves are superimposed based on their established operation control plan curves (i.e., the expected output curves before active regulation). Finally, by integrating the above two sets of data, a complete baseline load curve is generated. The calculation formula is as follows:

[0090] P 基线 =P 非可直控_预测 -P 分布式电源 +P 可直控_原 (3)

[0091] Among them, P 基线 This represents the final generated baseline load curve; P 非可直控_预测 P represents the non-controllable load component predicted by the TGDA model. 分布式电源 The raw output data of distributed power sources extracted during the pre-separation processing stage; P 可直控_原 This refers to the directly controllable load curve under the premise of no active regulation.

[0092] The method described in this invention will be verified through specific experiments below.

[0093] In this experiment, an actual integrated source-grid-load-storage park in Shandong Province was selected as a case study for algorithm testing. The park configuration is shown in Table 1 below:

[0094] Table 1 Installed Capacity of Various Resources in the Park

[0095]

[0096]

[0097] In this experiment, a complete dataset was constructed using historical data from the past 360 days, with the first 300 days used for model training and the last 60 days as the test set. The method of this invention was compared with the historical mean method, the LSTM model, and the Transformer model. Evaluation metrics included MAPE (Mean Absolute Percentage Error) and R-squared value. 2 (Coefficient of determination) and NRMSE (Normalized Root Mean Square Error). The evaluation calculation formula is as follows:

[0098]

[0099] Where n represents the total number of points in the sequence, y i Represents the actual value. Represents the value to be evaluated. y represents the average of the actual values. max y represents the maximum value among the actual values. min This represents the minimum value among the actual values.

[0100] In this experiment, the method described in this invention was used to perform deep extraction of front-end temporal features on the dataset using a TCN network incorporating an attention mechanism. The obtained temporal features were then input into a GRU layer with an attention mechanism to obtain the prediction curve. The comparison results with other models are shown in Table 2 below:

[0101] Table 2 Comparison of prediction results for each model

[0102] Model MAPE (%) [R 2 ]] NRMSE Historical average method 17.96 0.33796 0.18429 LSTM 11.595 0.67338 0.12944 Transformer 6.8472 0.89678 0.072768 The present invention 6.2836 0.91293 0.066832

[0103] The test results show that, over a 60-day average, this invention outperforms current mainstream deep learning models in all three key metrics. Compared to the Transformer model, the MAPE value is reduced by more than 8%; compared to the LSTM model, the MAPE value is improved by more than 30%. Overall, whether in terms of MAPE, R... 2 In terms of NRMSE metrics, the method proposed in this invention achieves the best performance, fully demonstrating its significant advantages in prediction accuracy, model robustness, and generalization ability.

[0104] In this experiment, after generating the prediction curve of non-controllable load using the method described in this invention, the previously pre-separated output curves of distributed new energy power stations are superimposed and restored according to the original data. For controllable loads, the curves are superimposed according to their predetermined operation control plan curves (i.e., the expected output curves before active regulation is implemented) to generate the final baseline load curve.

[0105] Furthermore, to fully demonstrate the versatility of this method, baseline load curves were randomly selected from the park under three different shift systems (three-shift, two-shift, and one-shift), in different seasons, and under different weather conditions (sunny and rainy days). The curves were then plotted as follows: Figures 4 to 12 As shown:

[0106] Figures 4 to 6 The diagram illustrates the baseline load curves obtained by the method of this invention, along with those obtained by the historical mean method, LSTM model, and Transformer model, under three different shift schedules. As the results show, the baseline load curve generated by this invention exhibits superior fitting performance compared to other existing methods under various production modes. Whether during periods of significant load fluctuation or relatively stable load phases, this invention effectively captures the actual load change trend, demonstrating strong adaptability and robustness. This advantage is particularly evident in complex and ever-changing electricity consumption scenarios, indicating that this invention not only operates stably under ideal conditions but also maintains high accuracy and reliability in real-world applications. Comparative analysis with other mainstream methods demonstrates that this invention significantly improves the accuracy of baseline load prediction, providing stronger data support and technical assurance for subsequent energy management, power dispatching, and demand response strategy development.

[0107] Figures 7 to 10 The paper presents baseline load curves obtained by the method of this invention, along with those obtained by the historical mean method, LSTM model, and Transformer model, under four seasons. Analysis of the baseline load generation effect across the four seasons shows that the method proposed in this invention exhibits the best fitting performance in all four seasons. Furthermore, this invention can effectively track and reflect the differences in load amplitude between different seasons, fully verifying its superior performance and wide applicability.

[0108] Figures 11 to 12 The figure shows the baseline load curves obtained by the proposed method under sunny and rainy conditions, compared with those obtained by the historical mean method, LSTM model, and Transformer model. As can be seen from the figure, the proposed method exhibits the best fitting results under both sunny and rainy conditions. Under sunny conditions, due to sufficient sunlight and stable temperature, user electricity consumption behavior is relatively regular, resulting in a relatively stable baseline load trend. However, under rainy conditions, load fluctuations are significantly amplified by factors such as sudden weather changes, temperature drops, and changes in user activity patterns. Even so, the proposed method can still accurately capture load change trends and effectively suppress prediction interference caused by sudden weather changes. This fully demonstrates the strong resistance of the proposed method to baseline load prediction interference under variable weather conditions, exhibiting good robustness and adaptability.

[0109] The baseline load precise generation system for the integrated source-grid-load-storage park described in this invention includes:

[0110] The power curve pre-separation unit is used to obtain the gate power curve of the park and pre-separate the gate power curve according to the directly controllable load, the non-directly controllable load and the distributed power source to obtain the non-directly controllable load curve and the distributed power source output curve.

[0111] The dataset creation unit is used to create a dataset based on the non-controllable load curve and external characteristics, and to reconstruct the dataset according to time to obtain a reconstructed dataset; the external characteristics are obtained by screening relevant feature variables using the mutual information method;

[0112] The TGDA modeling unit is used to train and improve the TGDA model using the reconstructed dataset, and to obtain the predicted non-controllable load curve using the improved TGDA model; the feature mining layer of the improved TGDA model embeds an attention mechanism module after the TCN network structure.

[0113] The baseline load curve generation unit is used to overlay the predicted non-controllable load curve, the distributed power output curve, and the expected controllable load curve to obtain a complete baseline load curve.

[0114] The electronic device of the present invention includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the computer program is loaded onto the processor, it implements the baseline load accurate generation method for the integrated source-grid-load-storage park.

[0115] The computer program product of the present invention includes a computer program that, when executed by a processor, implements a method for accurately generating the baseline load of the integrated source-grid-load-storage park.

[0116] The computer-readable storage medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage devices, magnetic disk storage devices or other magnetic storage devices, flash memory or any other medium that can be used to store program code in the form of instructions or data structures and is accessible by a computer.

[0117] The processor is used to execute a computer program stored in memory to implement the various steps in the methods described in the above embodiments.

Claims

1. A method for accurately generating baseline load in an integrated source-grid-load-storage park, characterized in that, Includes the following steps: Obtain the gate power curve of the park, and perform pre-separation processing on the gate power curve according to the directly controllable load, the non-directly controllable load and the distributed power source to obtain the non-directly controllable load curve and the distributed power source output curve. A dataset was established based on the non-controllable load curve and external characteristics, and the dataset was reconstructed according to time to obtain the reconstructed dataset; the external characteristics were obtained by screening relevant characteristic variables using the mutual information method. An improved TGDA model was trained using a reconstructed dataset, and the improved TGDA model was used to obtain the predicted non-controllable load curve; the feature mining layer of the improved TGDA model embeds an attention mechanism module after the TCN network structure. The predicted non-controllable load curve, the distributed power generation output curve, and the expected controllable load curve are superimposed to obtain a complete baseline load curve.

2. The method for accurately generating baseline load in an integrated source-grid-load-storage park according to claim 1, characterized in that, The data reconstruction of the dataset according to time includes: The historical load data of several days before the generation date of the power curve at the gateway are concatenated with the corresponding external features on the generation date in chronological order and flattened into a one-dimensional feature vector, which serves as one dimension of the reconstructed dataset. Based on the one-dimensional feature vector, a sliding window with a step size of one day is used to continuously divide the data to obtain another dimension of the reconstructed dataset.

3. The method for accurately generating baseline load in an integrated source-grid-load-storage park according to claim 1, characterized in that, The improved TGDA model also includes: embedding an attention mechanism module between the two GRU layers to establish the temporal correlation of the improved TGDA model.

4. The method for accurately generating baseline load in an integrated source-grid-load-storage park according to claim 1, characterized in that, Complete baseline load curve P 基线 The calculation method is as follows: P 基线 =P 非可直控_预测 -P 分布式电源 +P 可直控_原 ; Among them, P 非可直控_预测 To predict non-controllable load curves; P 分布式电源 The output curve of the distributed power source obtained from the pre-separation process; P 可直控_原 This represents the expected controllable load curve when no active regulation is implemented.

5. The method for accurately generating baseline load in an integrated source-grid-load-storage park according to claim 1, characterized in that, The directly controllable loads include load devices whose power output can be freely adjusted at any time period, while the non-directly controllable loads include fixed loads that cannot be adjusted and load devices that can be adjusted to a limited extent.

6. A baseline load precise generation system for an integrated power generation, grid, load, and storage park, characterized in that, include: The power curve pre-separation unit is used to obtain the gate power curve of the park and pre-separate the gate power curve according to the directly controllable load, the non-directly controllable load and the distributed power source to obtain the non-directly controllable load curve and the distributed power source output curve. The dataset creation unit is used to create a dataset based on the uncontrollable load curve and external characteristics, and to reconstruct the dataset according to time to obtain a reconstructed dataset; among which... Exogenous features are obtained by screening relevant feature variables using mutual information. The TGDA modeling unit is used to train and improve the TGDA model using the reconstructed dataset, and to obtain the predicted non-controllable load curve using the improved TGDA model; the feature mining layer of the improved TGDA model embeds an attention mechanism module after the TCN network structure. The baseline load curve generation unit is used to overlay the predicted non-controllable load curve, the distributed power output curve, and the expected controllable load curve to obtain a complete baseline load curve.

7. The baseline load precise generation system for the integrated source-grid-load-storage park according to claim 6, characterized in that, In the dataset creation unit, the data reconstruction of the dataset according to time includes: The historical load data of several days before the generation date of the power curve at the gateway are concatenated with the corresponding external features on the generation date in chronological order and flattened into a one-dimensional feature vector, which serves as one dimension of the reconstructed dataset. Based on the one-dimensional feature vector, a sliding window with a step size of one day is used to continuously divide the data to obtain another dimension of the reconstructed dataset.

8. The baseline load precise generation system for the integrated source-grid-load-storage park according to claim 6, characterized in that, In the TGDA model modeling unit, the improved TGDA model also includes: the improved TGDA model temporal correlation establishment layer embeds an attention mechanism module between the two GRU layers.

9. The baseline load precise generation system for an integrated source-grid-load-storage park according to claim 6, characterized in that, In the baseline load curve generation unit, the complete baseline load curve P 基线 The calculation method is as follows: P 基线 =P 非可直控_预测 -P 分布式电源 +P 可直控_原 ; Among them, P 非可直控_预测 To predict non-controllable load curves; P 分布式电源 The output curve of the distributed power source obtained from the pre-separation process; P 可直控_原 This represents the expected controllable load curve when no active regulation is implemented.

10. The baseline load precise generation system for the integrated source-grid-load-storage park according to claim 6, characterized in that, In the power curve pre-separation unit, the directly controllable load includes load devices whose power output can be freely adjusted at any time period, and the non-directly controllable load includes non-adjustable fixed loads and load devices with limited adjustment.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the method for accurately generating the baseline load of the integrated source-grid-load-storage park according to any one of claims 1-5.

12. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for accurately generating the baseline load of the integrated source-grid-load-storage park according to any one of claims 1-5.

Images