Dynamic load regulation potential prediction method based on large generative model

Abstract

The invention discloses a dynamic load regulation potential prediction method based on a generative large model, and relates to the field of power load prediction and optimization. The method comprises four main steps of data collection and preprocessing, model construction and training, model reasoning and evaluation, and strategy optimization and decision making. A future load sequence is predicted through multi-modal data fusion and a generative large model, a load adjustment strategy is dynamically adjusted, and the optimization objectives of peak load shifting and power grid demand response are achieved. The problems that in a traditional load prediction method, the multi-modal data integration capacity is insufficient, a complete future scene cannot be generated, and intelligent optimization is insufficient are solved, and the method has the advantages of being high in precision, diversified, capable of achieving closed-loop optimization and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of power system load forecasting and dynamic regulation, and particularly to a method for predicting dynamic load regulation potential based on a generative large model. This method realizes high-precision prediction and dynamic regulation of future loads through multi-modal data fusion, generative modeling, and closed-loop optimization. Background Art

[0002] With the development of digital information technology, power grid enterprises have continuously promoted the construction, operation, and production technology upgrading of intelligent measurement and control general terminals. With the overall goal of "improving quality and increasing efficiency", they carry out the intelligent and integrated construction of intelligent measurement and control general terminals, giving full play to the role of production factors of data in the substation field, and strongly supporting the upgrading of digital power grids, digital enterprises, digital services, and digital industries. Among them, intelligent measurement and control general terminals can include various power grid sensors.

[0003] In the traditional method, intelligent measurement and control general terminals are divided according to professional modules, and each module separately collects and processes data, resulting in prominent information island problems in intelligent measurement and control general terminals and relatively low intelligent level of data processing in intelligent measurement and control general terminals.

[0004] Moreover, the existing networking methods for intelligent measurement and control general terminals are relatively single, basically fixed networking methods. The number and communication distance of intelligent terminals in the networking structure remain unchanged. However, when facing the situation where the communication volume exceeds the information volume that can be transmitted by intelligent terminals in the networking structure, either information transmission omissions occur, or when re-networking, the structure needs to be adjusted extensively, consuming a large amount of manpower and material resources;

[0005] During the information transmission process of intelligent measurement and control general terminals, due to different data types, there are a large number of noise data in the finally aggregated information. Existing technologies either do not perform information filtering, or the filtering method is inefficient, and even filter some non-noise data, affecting the final data fusion result. Summary of the Invention

[0006] In view of the problems existing in the prior art, the present invention is proposed.

[0007] Therefore, the problems to be solved by the present invention include insufficient ability to integrate complex multi-modal data, limited ability to capture dynamic changes, and lack of intelligent optimization ability.

[0008] To solve the above technical problems, the present invention provides the following technical solutions:

[0009] In a first aspect, an embodiment of the present invention provides a method for predicting dynamic load regulation potential based on a generative large model, which includes collecting multimodal data for data filling processing; fusing the multimodal data through the generative large model; generating future load predictions in combination with historical data, and realizing closed-loop optimization of the load regulation strategy by using the prediction results and the feedback mechanism.

[0010] As a preferred solution of the method for predicting dynamic load regulation potential based on a generative large model according to the present invention, wherein: the collecting multimodal data for data filling processing includes data collection and preprocessing: for collecting historical load data, meteorological data, user behavior data, and market price data, and performing normalization, time alignment, and missing value filling processing;

[0011] The data collection and preprocessing specifically include:

[0012] Using a normalization formula to standardize various input data into a unified range to reduce the impact of dimension differences on model training. The normalization formula is as follows.

[0013]

[0014] Wherein, x(t) represents the original time series data, and min(x) and max(x) are the minimum and maximum values of the data respectively.

[0015] Align different modality data to a unified time granularity, and the time granularity can be at the minute, hour, or day level;

[0016] Using a deep learning method to complete the missing values, specifically expressed as:

[0017]

[0018] Wherein, f AE is an autoencoder model, and x context is the context data within the time window.

[0019] As a preferred solution of the method for predicting dynamic load regulation potential based on a generative large model according to the present invention, wherein: the fusing the multimodal data through the generative large model includes model construction and training: capturing multimodal feature interactions and time dependencies through an embedding layer and an encoding layer;

[0020] The model construction and training specifically include:

[0021] Mapping multimodal inputs into a unified feature space through an embedding matrix:

[0022] Embedding(x t ) = E·x t + b

[0023] Among them, E is the embedding matrix, b is the bias vector, and x t is the multi-modal input feature vector;

[0024] The multi-head attention mechanism based on Transformer is adopted to capture the time series dependence and the inter-modal interaction characteristics:

[0025] Z l = MultiHeadAttention(Q l , K l , V l ) + FeedForward(Z l-1 )

[0026] Among them, Q, K, and V respectively represent the query, key, and value vectors, and Z l is the feature representation of the l-th layer;

[0027] The model generates the future load sequence in an autoregressive manner:

[0028]

[0029] Among them, is the predicted load at the next moment, and θ is the model parameter.

[0030] As a preferred solution of the dynamic load regulation potential prediction method based on the generative large model described in the present invention, wherein: the loss function of the model includes:

[0031] The mean square error (MSE) is used to evaluate the difference between the predicted value and the true value:

[0032]

[0033] Among them, is the predicted value, and x t is the true value;

[0034] The model parameters are constrained by L2 regularization:

[0035]

[0036] Among them, λ is the regularization coefficient;

[0037] Total loss: Considering both the prediction accuracy and the model complexity:

[0038] L = L MSE + L reg

[0039] Among them, L MSE is the model stability, and L regis the total loss.

[0040] As a preferred embodiment of the method for predicting the dynamic load regulation potential based on a generative large model according to the present invention, wherein: generating a future load prediction by combining historical data, and realizing the closed-loop optimization of the load regulation strategy by using the prediction result and the feedback mechanism includes model inference and evaluation, generating a future load sequence based on the input historical data window, and evaluating the prediction accuracy and correlation;

[0041] The model inference and evaluation specifically include:

[0042] Generating a future load sequence based on the historical data window {x t-w ,..., x t}:

[0043]

[0044] where w is the window length and h is the prediction step;

[0045] Evaluating the model performance by the mean absolute percentage error (MAPE) and the Pearson correlation coefficient:

[0046] Mean absolute percentage error:

[0047]

[0048] Pearson correlation coefficient:

[0049]

[0050] As a preferred embodiment of the method for predicting the dynamic load regulation potential based on a generative large model according to the present invention, wherein: generating a future load prediction by combining historical data, and realizing the closed-loop optimization of the load regulation strategy by using the prediction result and the feedback mechanism further includes strategy optimization and decision-making, optimizing the load regulation strategy by using the prediction result, and realizing closed-loop optimization through feedback learning.

[0051] The strategy optimization and decision-making specifically include:

[0052] Optimizing the load regulation strategy based on the predicted future load sequence and constructing an objective function:

[0053]

[0054] where C 调节 (x t ) represents the cost of load regulation;

[0055] Using the reinforcement learning method to realize strategy optimization, specifically:

[0056]

[0057] Among them, Q(s t , a t ) represents the value function of the current state s t and the action a t , r t is the immediate reward, and γ is the discount factor.

[0058] As a preferred solution of the method for predicting the dynamic load regulation potential based on a generative large model according to the present invention, wherein: the feedback learning updates the parameters of the generative large model by means of reinforcement learning or online training, specifically including:

[0059] Readjust the model weights based on the latest load data;

[0060] Evaluate the actual effect of the regulation strategy to generate a reinforcement learning reward signal;

[0061] This method is applicable to the optimization scheduling scenarios of peak shaving and valley filling, real-time demand response, and new energy access, and improves the operation efficiency and economy of the power system through a closed-loop optimization process.

[0062] In a second aspect, an embodiment of the present invention provides a system for sequentially obtaining the inertia of multiple resources and primary frequency regulation ancillary services, which includes a data collection and preprocessing unit that collects historical load data, meteorological data, user behavior data, and market price data, and performs normalization, time alignment, and missing value filling processing;

[0063] A model construction and training unit that uses a generative large model to fuse multimodal data, and captures multimodal feature interactions and time dependencies through an embedding layer and an encoding layer;

[0064] A model inference and evaluation unit that generates future load sequences based on the input historical data window, and evaluates the prediction accuracy and correlation;

[0065] A strategy optimization and decision-making unit that uses the prediction results to optimize the load regulation strategy and achieves closed-loop optimization through feedback learning.

[0066] In a third aspect, an embodiment of the present invention provides a computer device, including a memory and a processor, wherein: when the computer program instructions are executed by the processor, the steps of a method for predicting the dynamic load regulation potential based on a generative large model as described in the first aspect of the present invention are implemented.

[0067] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, wherein: when the computer program instructions are executed by the processor, the steps of a method for predicting the dynamic load regulation potential based on a generative large model as described in the first aspect of the present invention are implemented.

[0068] The beneficial effects of the present invention are as follows: high-precision prediction of future complete load sequences is achieved through generative large models, overcoming the deficiency of traditional methods that can only perform single-point or short-term predictions; through multi-modal data fusion, the adaptability to complex power load environments is significantly improved; and the closed-loop optimization based on feedback learning enables a positive interaction between prediction and regulation strategies, continuously enhancing the regulation efficiency and intelligence level of the system. Therefore, the present invention can demonstrate higher accuracy, flexibility, and economy in the load regulation of actual power systems, providing a comprehensive solution for peak shaving and valley filling, demand-side response, and the effective access of new energy. This method has prominent advantages such as high precision, diversification, and closed-loop optimization, and is applicable to the load prediction and scheduling requirements in various complex scenarios of modern power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0070] Figure 1 It is a flowchart of a method for predicting the dynamic load regulation potential based on a generative large model;

[0071] Figure 2 It is a flowchart of a load time series prediction method based on the Transformer model;

[0072] Figure 3 It is a structural diagram of a load regulation strategy based on the Q-learning reinforcement learning algorithm. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0073] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention in conjunction with the drawings in the specification.

[0074] Many specific details are set forth in the following description in order to provide a thorough understanding of the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0075] Secondly, the "one embodiment" or "embodiment" mentioned herein refers to specific features, structures or characteristics that may be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not all refer to the same embodiment, nor is it an embodiment that is separate or selectively mutually exclusive with other embodiments.

[0076] Embodiment 1

[0077] Referring to Figures 1 - 3 , which is the first embodiment of the present invention. This embodiment provides a method for predicting the dynamic load regulation potential based on a generative large model, including:

[0078] S1. Data collection and preprocessing; specifically, first collect multimodal data from different sources. Historical load data is collected through smart meters, including user electricity consumption records and distributed energy output records at different time intervals. Meteorological data comes from monitoring stations or online services, including environmental information such as temperature, humidity, wind speed, and light intensity. User behavior data is recorded by smart home devices, reflecting user electricity consumption habits, such as the usage frequency and time distribution of devices. Market price data is collected from the electricity market, mainly including electricity price fluctuations and peak-valley period information.

[0079] In order to unify the data format, preprocessing operations need to be carried out. First is data normalization, which maps data with different dimensions to a unified range to reduce interference with model training:

[0080]

[0081] where x(t) is the original data, and min(x) and max(x) are the minimum and maximum values of the data respectively.

[0082] Next, perform time alignment operations on the multimodal data to align various types of data at a unified time granularity, such as hourly or minute granularity. The specific operations are as follows:

[0083] The first step is to define a unified set of time indices, set the time step as Δ (such as 1 hour, 15 minutes, or 5 minutes, etc.), and divide a series of equally spaced time points within a given time range:

[0084]

[0085] where adjacent time points satisfy t j+1 -t j = Δ.

[0086] The second step is to align each modality of data. Assume that the data of a certain modality is originally sampled at irregular or different intervals {s 1 , s 2 , …, sM}, and the observed values corresponding to each moment are {x(s 1 ), x(s 2 ), …, x(s M )}. In order to map it to a unified time index the following interval averaging method is adopted:

[0087]

[0088] where x′(t j ) represents the alignment result of the modal data in the time interval [t j , t j + Δ); if there are no observed values in the interval, missing value handling or interpolation strategies can be considered, such as linear interpolation or filling with an autoencoder.

[0089] In the third step, the multi-modal data is merged, and the above alignment process is performed on each modality such as meteorological data, user behavior data, market price data, etc., so as to obtain the feature sequence corresponding to {t 1 , t 2 , …, t N}. In this way, the alignment result of the multi-modal data at a unified time step can be formed, providing consistent input for subsequent model training or prediction.

[0090] For missing data, a deep learning autoencoder is used for filling:

[0091]

[0092] where f AE is the autoencoder model, and the input is the context data x context within the time window. This can effectively fill the data gaps and ensure the accuracy of subsequent analysis.

[0093] It should be noted that through the normalization operation, the differences between different dimensions are eliminated, avoiding the adverse effects of too large or too small data scales on model training;

[0094] Time alignment enables the correlation analysis of meteorology, market price, user behavior and historical load data under the same time benchmark, improving the model's ability to capture the correlation of multi-modal features;

[0095] The deep learning autoencoder is used for missing value filling, which can more finely mine the internal patterns and context information of the time series, significantly reducing the impact of missing data on the model accuracy and stability.

[0096] S2. Model construction and training; specifically, after data preprocessing, the multimodal data is input into the generative large model for modeling and training. In the present invention, the Transformer model is adopted. Please refer to Figure 2 , and the Transformer model consists of an embedding layer, an encoder layer, and a decoder layer. The functions of each layer are as follows.

[0097] (1) Embedding layer

[0098] The multimodal data (such as historical load, meteorology, user behavior, market price, etc.) will first be processed by the embedding layer and combined with positional encoding to form an input sequence. The multimodal data is mapped to a unified feature space through an embedding matrix, expressed as:

[0099] Embedding(x t ) = E·x t + b,

[0100] where E is the embedding matrix, x t is the input feature vector at time step t, and b is the bias term. The role of this layer is to eliminate the differences between modalities and represent the data in a unified space.

[0101] (2) Encoder layer

[0102] The encoding layer utilizes the multi-head self-attention sublayer and the feed-forward network sublayer, and cooperates with operations such as residual connection and layer normalization between each sublayer. The specific calculation is as follows:

[0103] Z l = MultiHeadAttention(Q l , K l , V l ) + FeedForward(Z l-1 ),

[0104] where Q l , K l , V l are the query, key, and value vectors respectively, representing the features of different time steps or modalities. The multi-head attention mechanism allows the model to focus on different patterns in the time series (such as long-term trends or short-term fluctuations).

[0105] The calculation formula for the attention weight is

[0106]

[0107] where d kis the dimension of the key vector.

[0108] (3) Decoder layer

[0109] Based on the multi-head self-attention sub-layer, a multi-head cross-attention sub-layer (Multi-Head Cross-Attention) that interacts with the encoder output is also added. Then, it cooperates with the feed-forward network sub-layer for information processing, still with residual connections and layer normalization. The model generates future load time series through the decoding layer. Based on the historical time window {x t-w ,..., x t}, the decoding layer generates future load values in an autoregressive manner:

[0110]

[0111] where θ is the model parameter and f is the decoding function. The decoding layer can gradually predict the complete future load sequence.

[0112] (4) Loss function

[0113] The goal of model training is to minimize the prediction error and the regularization term:

[0114] L = L MSE + L reg ,

[0115] where the prediction error L MSE is defined as:

[0116]

[0117] The regularization term L reg is used to limit the model complexity:

[0118]

[0119] where λ is the regularization coefficient.

[0120] It should be noted that the embedding layer unifies the representation space of multi-modal features, avoids the problem of feature distribution differences when directly modeling on the original heterogeneous data, and is beneficial for the model to better aggregate information from each modality; the multi-head attention mechanism can simultaneously focus on long-term and short-term dependencies as well as the interaction information between multi-modalities, significantly improving the ability to capture complex load dynamic changes;

[0121] Autoregressive generation can provide a complete load prediction sequence for future time periods, rather than just a single-point prediction, which helps subsequent strategy optimization to make more flexible decisions based on the overall trend. MSE measures the deviation between the predicted load and the actual load, guiding the model to continuously improve the prediction accuracy during training; L2 regularization effectively suppresses the over-complication of the model, enhancing the generalization ability of the model and its robustness to noise; comprehensively incorporating the prediction accuracy and model complexity into the objective function can effectively avoid overfitting while pursuing high accuracy, improving the practicality of the model in multi-modal complex scenarios.

[0122] S3. Model Inference and Evaluation; Specifically, the trained generative large model is used for predicting future load sequences. Input the historical time window {x t-w ,..., x t}, and the model generates a prediction sequence:

[0123]

[0124] where h is the prediction step.

[0125] To verify the model performance, the following evaluation metrics are adopted:

[0126] (1) Mean Absolute Percentage Error (MAPE)

[0127] It is used to measure the relative error between the predicted value and the actual value:

[0128]

[0129] The closer the MAPE value is to 0, the more accurate the prediction result; when MAPE≈0%, it means the prediction result is very accurate; when MAPE≤5%, it is generally regarded as high-precision prediction; when 5% < MAPE≤10%, it indicates that there are certain errors in the prediction but still within an acceptable range; if MAPE>10%, it means that the model needs to be further optimized or data features need to be added to improve the prediction accuracy.

[0130] (2) Pearson Correlation Coefficient

[0131] It evaluates the linear correlation between the predicted value and the actual value:

[0132]

[0133] where and are the means of the actual value and the predicted value respectively.

[0134] In the Pearson correlation coefficient, the value range of r is [-1, 1]. When r is greater than 0, it indicates a positive correlation; when r is less than 0, it indicates a negative correlation; when r = 0, it indicates no linear correlation between the two;

[0135] Generally speaking, |r| → 1 indicates a stronger linear correlation; when r > 0.8, it can be considered that the predicted value and the true value have a strong positive consistency; if r < 0, it means that the prediction result is negatively correlated with the true value, and the model training and data processing process need to be further adjusted or checked.

[0136] S4. Strategy optimization and decision-making; specifically, based on the prediction results, a load regulation strategy is constructed with the goal of minimizing the load regulation cost:

[0137]

[0138] where C 调节 (xt) represents the cost when adjusting the load x t . The cost function can be defined as a linear or non-linear form according to the actual situation.

[0139] To further optimize the regulation strategy, the present invention adopts the Q-learning reinforcement learning algorithm. Please refer to Figure 3 , the Agent of reinforcement learning takes the predicted load output by the Transformer as part of the environmental state, and selects the optimal load regulation action (a) based on this. In each step of interaction, the environment feedbacks the immediate reward (r) and updates it to the new state (s'). After iterative loops, the Agent learns the optimal regulation strategy that can balance peak shaving and valley filling and economy under various operating scenarios. The working mechanism of Q-learning is as follows:

[0140] State (s): Represents the current load and predicted load sequence, including the Transformer predicted load, market price, etc.

[0141] Action (a): Represents the load regulation strategy, such as specific operations like turning on / off certain adjustable devices, adjusting the output, switching between peak and valley, etc.

[0142] Immediate reward (r): Designed according to "electricity cost", "peak shaving and valley filling effect", "degree of meeting demand", etc.

[0143] Q function update: Continuously iteratively update the Q function or Q network through .

[0144] The goal of maximizing the cumulative reward can be expressed as the following optimal action-value function, that is, the optimal Q-learning equation:

[0145]

[0146] where r tis the immediate reward, and γ is the discount factor. Note that if the system is a Markov process and the current state is only related to the previous state of the system, then the optimal policy for the current state must be the policy that maximizes the cumulative reward in the current state, that is

[0147]

[0148] It is not difficult to find that for any non-optimal policy π, if its policy is changed to π* and Q(s, π*(s)) ≥ V(s, π),

[0149] then its value function is monotonically increasing for policy improvement. Therefore, for the current policy π and state s, its policy improvement is equivalent to the following optimization:

[0150]

[0151]

[0152] It should be noted that directly associating the high-precision prediction results with the adjustment strategy can respond more timely and accurately to the changes in the load demand of the power system; the introduction of reinforcement learning enables the adjustment strategy to be continuously optimized through continuous exploration and feedback learning in different operating scenarios, forming a more adaptable closed-loop decision-making; it has good application prospects in scenarios such as peak shaving and valley filling, real-time demand response, and new energy access, improving the economic benefits and operating efficiency of the entire power system.

[0153] Furthermore, this embodiment also provides a system for predicting the dynamic load regulation potential based on a generative large model, including:

[0154] A data collection and preprocessing unit that collects historical load data, meteorological data, user behavior data, and market price data, and performs normalization, time alignment, and missing value filling processing;

[0155] A model construction and training unit that uses a generative large model to fuse multi-modal data, and captures multi-modal feature interactions and time dependencies through an embedding layer and an encoding layer;

[0156] A model inference and evaluation unit that generates future load sequences based on the input historical data window, and evaluates the prediction accuracy and correlation;

[0157] A policy optimization and decision-making unit that uses the prediction results to optimize the load regulation policy and realizes closed-loop optimization through feedback learning.

[0158] This embodiment also provides a computer device, which is applicable to the situation of the method for predicting the dynamic load regulation potential based on a generative large model, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for predicting the dynamic load regulation potential based on a generative large model as proposed in the above embodiment.

[0159] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad set on the outer shell of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0160] This embodiment also provides a storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the method for predicting the dynamic load regulation potential based on a generative large model as proposed in the above embodiment.

[0161] In summary, the present invention realizes high-precision prediction of future complete load sequences through a generative large model, overcoming the deficiency that traditional methods can only perform single-point or short-term predictions; through multi-modal data fusion, it greatly improves the adaptability to complex power load environments; and the closed-loop optimization based on feedback learning enables a virtuous interaction between prediction and regulation strategies, continuously improving the regulation efficiency and intelligence level of the system. Therefore, the present invention can demonstrate higher accuracy, flexibility, and economy in the load regulation of an actual power system, providing a comprehensive solution for peak shaving and valley filling, demand-side response, and the effective access of new energy. This method has prominent advantages such as high precision, diversification, and closed-loop optimization, and is applicable to the load prediction and scheduling requirements in various complex scenarios in modern power systems.

[0162] Embodiment 2

[0163] Referring to Figures 1 - 3 , this is the second embodiment of the present invention. This embodiment provides a method for predicting the dynamic load regulation potential based on a generative large model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0164] Taking the multimodal data collected from different sources within a day (with an hourly time granularity) in a certain area as an example, it covers historical load data, distributed energy output, meteorological data, user behavior data, market prices, etc. The specific data is shown in Table 1 as follows:

[0165] Table 1 Simulation Experiment Data Table

[0166]

[0167]

[0168] It should be noted that by comprehensively analyzing and utilizing multimodal data, including historical load, distributed energy output, meteorological conditions, user behavior, and market prices, etc., this method can achieve high-precision prediction and dynamic regulation of the power system load. This comprehensive analysis not only improves the accuracy and reliability of the prediction but also optimizes the access of new energy and user electricity consumption behavior, thus enhancing the operation efficiency and economy of the power system.

[0169] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting dynamic load regulation potential based on a generative large model, characterized by: include, Collect multimodal data for data filling processing; Fusion of multimodal data through generative large models; Combine historical data to generate future load forecasts, and use the forecast results and feedback mechanism to achieve closed-loop optimization of load regulation strategies.

2. A method for predicting dynamic load regulation potential based on a generative large model as claimed in claim 1, characterized in that: The data filling process of collecting multimodal data includes data collection and preprocessing: collecting historical load data, meteorological data, user behavior data and market price data, and performing normalization, time alignment and missing value filling processes; The data collection and preprocessing specifically include: A normalization formula is used to standardize various input data to a unified range to reduce the impact of dimensional differences on model training. The normalization formula is as follows: Among them, x(t) represents the original time series data, min(x) and max(x) are the minimum and maximum values ​​of the data respectively. Align different modal data to a unified time granularity, which can be minutes, hours, or days; The deep learning method is used to complete the missing values, which is specifically expressed as follows: Among them, f AE is the autoencoder model, x context is the context data within the time window.

3. A method for predicting dynamic load regulation potential based on a generative large model as claimed in claim 2, characterized in that: The fusing of multimodal data through a generative large model includes model construction and training: capturing multimodal feature interactions and time dependencies through an embedding layer and an encoding layer; The model construction and training specifically include: The multimodal input is mapped into a unified feature space through the embedding matrix: Embedding(x t )=E·x t +b Among them, E is the embedding matrix, b is the bias vector, and x t Input feature vector for multimodal; The Transformer-based multi-head attention mechanism is used to capture time series dependencies and inter-modal interaction characteristics: From l =MultiHeadAttention(Q l ,K l ,V l )+FeedForward(Z l-1 ) Among them, Q, K, V represent query, key and value vectors respectively, and Z l is the feature representation of the lth layer; The model generates future load series by autoregression: in, is the predicted load at the next moment, and θ is the model parameter.

4. A method for predicting dynamic load regulation potential based on a generative large model as claimed in claim 3, characterized in that: The loss function of the model includes: The mean square error (MSE) is used to evaluate the difference between the predicted value and the true value: in, is the predicted value, x t is the true value; Constrain the model parameters through L2 regularization: Among them, λ is the regularization coefficient; Considering prediction accuracy and model complexity: L=L MSE +L reg Where L MSE is the model stability, L reg is the total loss.

5. A method for predicting dynamic load regulation potential based on a generative large model as claimed in claim 4, characterized in that: The method of combining historical data to generate future load forecasts and using the forecast results and feedback mechanism to achieve closed-loop optimization of load regulation strategies includes model reasoning and evaluation: generating future load sequences based on input historical data windows and evaluating forecast accuracy and relevance; The model reasoning and evaluation specifically include: Based on the historical data window {x t-w , …, x t Generate future load sequence: Among them, w is the window length and h is the prediction step length; The model performance was evaluated by the mean absolute percentage error (MAPE) and Pearson correlation coefficient: Mean absolute percentage error: Pearson correlation coefficient: in and are the means of the actual and predicted values, respectively.

6. A method for predicting dynamic load regulation potential based on a generative large model as claimed in claim 5, characterized in that: The method of generating future load forecasts in combination with historical data and realizing closed-loop optimization of load regulation strategies by using forecast results and feedback mechanisms also includes strategy optimization and decision-making: optimizing load regulation strategies by using forecast results and realizing closed-loop optimization by feedback learning; The strategy optimization and decision-making specifically include: Optimize the load regulation strategy based on the predicted future load sequence and construct the objective function: Among them, C 调节 (x t ) represents the cost of load regulation; The reinforcement learning method is used to achieve strategy optimization, specifically: Among them, Q(s t , a t ) indicates the current state s t and action a t The value function, r t is the immediate reward, and γ is the discount factor.

7. A method for predicting dynamic load regulation potential based on a generative large model as claimed in claim 6, characterized in that: The feedback learning updates the parameters of the generative large model by means of reinforcement learning or online training, specifically including: Re-adjust model weights based on the latest load data; Evaluate the actual effect of the adjustment strategy and generate reinforcement learning reward signals; This method is suitable for optimized scheduling scenarios such as peak shaving and valley filling, real-time demand response and new energy access, and improves the operating efficiency and economy of the power system through a closed-loop optimization process.

8. A system for predicting dynamic load regulation potential based on a generative large model, based on a method for predicting dynamic load regulation potential based on a generative large model according to any one of claims 1 to 7, characterized in that: Also includes: The data collection and preprocessing unit collects historical load data, meteorological data, user behavior data, and market price data, and performs normalization, time alignment, and missing value filling processing; The model building and training unit uses a generative large model to fuse multimodal data and capture multimodal feature interactions and time dependencies through the embedding layer and encoding layer; Model reasoning and evaluation unit, which generates future load series based on the input historical data window and evaluates the prediction accuracy and relevance; The strategy optimization and decision-making unit uses the prediction results to optimize the load regulation strategy and achieves closed-loop optimization through feedback learning.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the processor implements the steps of a method for predicting dynamic load regulation potential based on a generative large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a method for predicting dynamic load regulation potential based on a generative large model as described in any one of claims 1 to 7 are implemented.

Images