Power grid load intelligent prediction method and device based on multi-model cooperation, equipment and medium

Abstract

The application relates to a power grid load intelligent prediction method and device based on multi-model cooperation, equipment and a medium, and relates to the technical field of power grid load prediction. The method comprises the following steps: acquiring first load correlation data corresponding to a power grid; processing the first load correlation data based on a first prediction model to obtain a first prediction result corresponding to the power grid; wherein the first prediction model is a data-driven model constructed based on a multi-layer bidirectional long short-term memory network; processing the first load correlation data and the first prediction result based on a second prediction model to obtain a second prediction result corresponding to the power grid; wherein the second prediction model is a physical information neural network model in which a physical law constraint of power grid operation is embedded as a regularization term in a model loss function; the physical law constraint is dynamically changed; and obtaining a third prediction result corresponding to the power grid based on the first prediction result and the second prediction result. The method can obtain an accurate load prediction result.

Description

Technical Field

[0001] This application relates to the field of power grid load forecasting technology, and in particular to a method, device, equipment and medium for intelligent power grid load forecasting based on multi-model collaboration. Background Technology

[0002] As a crucial component of power system dispatch optimization, energy allocation, and safe and stable operation, the accuracy and reliability of power grid load forecasting directly impact the economy and security of the power grid. In recent years, the continuous increase in the proportion of renewable energy grid connection and the increasingly diversified electricity consumption behavior of users have resulted in power grid load exhibiting strong nonlinearity, high volatility, and multi-scale characteristics. Traditional forecasting methods based on experience or simple statistical models are no longer sufficient to meet the needs of complex scenarios.

[0003] In existing technologies, power grid load forecasting methods are mainly divided into three categories: data-driven methods, physical mechanism methods, and hybrid methods. However, all of them suffer from low forecasting accuracy.

[0004] Therefore, there is an urgent need for a method that can accurately predict grid load. Summary of the Invention

[0005] This application provides a method, device, equipment, and medium for intelligent power grid load forecasting based on multi-model collaboration, which can obtain accurate load forecasting results.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] Firstly, this application provides a method for intelligent power grid load forecasting based on multi-model collaboration, including:

[0008] Obtain the first load-related data corresponding to the power grid;

[0009] Based on the first prediction model, the first load-related data is processed to obtain the first prediction result corresponding to the power grid; wherein, the first prediction model is a data-driven model built on a multi-layer bidirectional long short-term memory network.

[0010] Based on the second prediction model, the first load-related data and the first prediction result are processed to obtain the second prediction result corresponding to the power grid; wherein, the second prediction model is a physical information neural network model that embeds the physical law constraints of power grid operation as a regularization term into the model loss function; the physical law constraints are dynamically changing;

[0011] Based on the first and second prediction results, the third prediction result corresponding to the power grid is obtained.

[0012] In one embodiment, based on the first and second prediction results, a third prediction result corresponding to the power grid is obtained, including:

[0013] The first and second prediction results are concatenated according to time steps to form a fused feature vector;

[0014] The fused feature vector is input into the attention network to obtain the dynamic weights of the first and second prediction results at each time step.

[0015] Based on dynamic weights, the first and second prediction results are weighted and summed to obtain the third prediction result corresponding to the power grid.

[0016] In one embodiment, based on a first prediction model, the first load-related data is processed to obtain a first prediction result corresponding to the power grid, including:

[0017] Based on the variational mode decomposition algorithm, the load sequence in the historical load data is decomposed to obtain K eigenmode function components; where K is a preset positive integer.

[0018] Based on the K intrinsic mode function components and the non-load characteristic data in the first load correlation data, the first prediction result corresponding to the power grid is obtained.

[0019] In one embodiment, based on K intrinsic mode function components and non-load characteristic data from the first load correlation data, a first prediction result corresponding to the power grid is obtained, including:

[0020] Trend terms are extracted from the K intrinsic mode function components, and the effective components are retained after removing noise components.

[0021] The effective components are aligned with the non-load characteristic data in the first load-related data according to time steps and then concatenated to form K new input sequences.

[0022] K new input sequences are fed into K parallel sub-prediction models, each of which is a bidirectional long short-term memory network with independently optimized structural parameters in the first prediction model.

[0023] The K sub-prediction results are superimposed to obtain the first prediction result.

[0024] In one embodiment, the training process for the first prediction model and the second prediction model includes:

[0025] Based on time series cross-validation, historical load correlation data are divided into training set, validation set and test set;

[0026] The initial first prediction model is trained using an optimizer, and the initial second prediction model is trained in real time by introducing a loss term constrained by physical laws.

[0027] If the initial first prediction model achieves the first prediction accuracy index, then the second prediction model is obtained; and if the initial second prediction model achieves the second prediction accuracy index, then the second prediction model is obtained.

[0028] In one embodiment, the updating method for physical law constraints includes:

[0029] Real-time acquisition of power grid change data;

[0030] Based on the changing data, the constraint parameters are updated through power grid flow calculations;

[0031] The physical law constraints are updated based on the updated constraint parameters.

[0032] In one embodiment, the post-processing steps for the third prediction result include:

[0033] The third prediction result is smoothed using a moving average algorithm based on an adaptive window size.

[0034] Based on the safety boundary conditions of power grid operation, the smoothed third prediction result is corrected.

[0035] Secondly, this application provides a smart grid load forecasting device based on multi-model collaboration, comprising:

[0036] The acquisition module is used to acquire the first load-related data corresponding to the power grid;

[0037] The first prediction module is used to process the first load-related data based on the first prediction model to obtain the first prediction result corresponding to the power grid; wherein, the first prediction model is a data-driven model built on a multi-layer bidirectional long short-term memory network.

[0038] The second prediction module is used to process the first load-related data and the first prediction result based on the second prediction model to obtain the second prediction result corresponding to the power grid. The second prediction model is a physical information neural network model that embeds the physical law constraints of power grid operation as regularization terms into the model loss function. The physical law constraints are dynamically changing.

[0039] The third prediction module is used to obtain the third prediction result corresponding to the power grid based on the first and second prediction results.

[0040] Thirdly, this application provides a computing device, including a memory and a processor;

[0041] The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of the first aspects.

[0042] Fourthly, this application provides a computer-readable storage medium for storing a computer program for performing the method as described in any one of the first aspects.

[0043] Fifthly, this application provides a computer program product comprising one or more computer instructions, wherein when the computer instructions are executed by a computer, the computer performs the method as described in any one of the first aspects.

[0044] As can be seen from the above technical solution, this application has at least the following beneficial effects:

[0045] In this application, by acquiring the first load-related data corresponding to the power grid, a data foundation is provided for subsequent power grid forecasting. Furthermore, based on a first prediction model constructed from a multi-layer bidirectional long short-term memory network, the first load-related data is processed to obtain the first prediction result corresponding to the power grid, achieving preliminary power grid forecasting. Further, based on a second prediction model that embeds dynamically changing physical constraints as regularization terms into the model loss function, the first load-related data and the first prediction result are processed to obtain the second prediction result corresponding to the power grid, achieving adjustment of the first prediction result. Subsequently, based on the first and second prediction results, a third prediction result corresponding to the power grid can be obtained. This scheme, by introducing a first prediction model, provides a preliminary forecast of the power grid load. Furthermore, by introducing a second prediction model that embeds dynamically changing physical constraints as regularization terms into the model loss function, the dynamic changes of the power grid are considered, laying the foundation for accurate load forecasting. Finally, by combining the first and second prediction results, an accurate load forecast result is obtained.

[0046] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0047] Figure 1 This is an application environment diagram of a smart power grid load forecasting method based on multi-model collaboration provided in the embodiments of this application;

[0048] Figure 2 This is a flowchart illustrating a smart power grid load forecasting method based on multi-model collaboration provided in an embodiment of this application.

[0049] Figure 3 This is a structural block diagram of a smart power grid load forecasting device based on multi-model collaboration provided in an embodiment of this application;

[0050] Figure 4 This is an internal structural diagram of a computer device provided in the embodiments of the application. Detailed Implementation

[0051] The terms "first," "second," and "third," etc., used in this application specification and accompanying drawings are used to distinguish different objects, not to limit a specific order.

[0052] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0053] To make the technical solution of this application clearer and easier to understand, the application scenarios of the technical solution of this application are described below with reference to the accompanying drawings. Figure 1 As shown in the figure, this figure is an application environment diagram provided by an embodiment of this application.

[0054] In this application scenario, server 104 is a hardware or software system with data storage, computing and interaction capabilities. It is the data receiving end and processing end, responsible for acquiring and subsequently utilizing data. Server 104 performs relevant calculations based on the first load correlation data, determines the third prediction result corresponding to the power grid, and transmits it to terminal 102 through the communication network for relevant technical personnel to view.

[0055] To make the technical solution of this application clearer and easier to understand, the following describes a method for intelligent power grid load forecasting based on multi-model collaboration, using server 104 as the execution subject, in conjunction with the above application scenario. Figure 2 As shown in the figure, this is a flowchart illustrating a method for intelligent power grid load forecasting based on multi-model collaboration provided in an embodiment of this application.

[0056] S201. Obtain the first load association data corresponding to the power grid.

[0057] The power grid, or power network, is a power system composed of power generation, transmission, transformation, distribution, and consumption, responsible for delivering electricity from power plants to users. The primary load-related data includes electricity consumption data, power grid operation data, historical meteorological data, future weather forecasts, and auxiliary data. For example, electricity consumption data includes load power consumption, real-time load power, and the operating status of electrical equipment; power grid operation data includes line voltage, line current, substation operating status, load margin of surrounding power grids, and conditions for backup power activation; meteorological data includes temperature, rainfall, wind speed, and extreme weather warnings; and auxiliary data includes holiday schedules, event schedules, and equipment maintenance plans.

[0058] For example, the server needs to connect to multiple independent data sources, such as the internal power grid system, meteorological platform, and dispatch center. It then obtains scattered data from each data source through active retrieval (e.g., requesting the latest data at fixed intervals) or passive reception (actively pushing data when the data source is updated). Subsequently, the server performs format unification, deduplication, and correlation verification on the received scattered data, such as verifying the logical matching degree between high temperature weather and load growth. The scattered data is then integrated into a complete and usable first load correlation data set to support subsequent load monitoring and power supply guarantee decisions.

[0059] S202. Based on the first prediction model, the first load-related data is processed to obtain the first prediction result corresponding to the power grid.

[0060] The first prediction model is a data-driven model built on a multi-layer bidirectional long short-term memory network. The multi-layer bidirectional long short-term memory network is an artificial intelligence algorithm. Long short-term memory can capture long-term dependencies in data (such as the seasonal patterns of electricity load). Bidirectional means learning from the past to the future and from the future to the past simultaneously, thus improving prediction accuracy. Multi-layer means further optimizing the processing capability of complex data by stacking multi-layer network structures. The first prediction result is the prediction output of the first prediction model on the first load-related data. It can be the electricity load value or trend for a certain period of time in the future (such as 1 hour or 1 day), which can provide preliminary prediction basis at the data level.

[0061] One possible approach is to decompose the load sequence in historical load data based on the variational mode decomposition algorithm to obtain K intrinsic mode function components; and to obtain the first prediction result corresponding to the main grid based on the K intrinsic mode function components and the non-load characteristic data in the first load correlation data.

[0062] Where K is a preset positive integer; Variational Mode Decomposition (VMD) is a data preprocessing algorithm that can decompose complex historical load sequences (such as 24-hour load data) into multiple simple, stable intrinsic mode function (IMF) components, making it easier for subsequent models to learn different features (such as trend features and fluctuation features) more accurately; IMF components are the output of the VMD algorithm, and each component represents a specific frequency fluctuation or trend in the original load sequence. For example, some components reflect the daily peak electricity consumption pattern, while others reflect random noise. The K components together constitute all the information of the original data; Non-load feature data are other data in the first load correlation data besides load data (such as electricity consumption and power), including meteorological data (temperature, rainfall), time data (date, holidays), power grid operation data (line voltage), etc. These data will affect load changes and need to be included in the forecast.

[0063] For example, trend terms are extracted from the K intrinsic mode function components, and the effective components are retained after removing noise components; the effective components are aligned with the non-load feature data in the first load-related data by time step and concatenated to form K new input sequences; the K new input sequences are respectively input into K parallel sub-prediction models, which are bidirectional long short-term memory networks with independently optimized structural parameters in the first prediction model; the K sub-prediction results are superimposed to obtain the first prediction result.

[0064] Among them, the sub-prediction model is a parallel computing unit within the first prediction model, with a total of K sub-models. Each sub-model is a bidirectional long short-term memory network with independently optimized structure and parameters, and is respectively responsible for processing the input sequence after concatenating the K intrinsic mode function components and the unloaded feature data.

[0065] For example, historical load data is processed using a variational mode decomposition algorithm. The originally complex load sequence (such as a year's daily electricity consumption data) is decomposed into K intrinsic mode function components, each corresponding to a specific load change characteristic, such as trend or fluctuation. Then, the K decomposed intrinsic mode function components are processed to extract the trend term reflecting the true load pattern, remove meaningless noise components, and retain only the effective components that can be used for prediction, reducing interference. Furthermore, the filtered effective components are then compared with non-load characteristic data (such as temperature and holidays) in the first load correlation data in a timely manner. The load components are aligned step-by-step (e.g., ensuring a one-to-one correspondence between the load components and the temperature data at that moment), and then concatenated into K new input sequences, each corresponding to a valid component. The K input sequences are then fed into K parallel sub-prediction models within the first prediction model. Each sub-model is a bidirectional long short-term memory network that independently optimizes parameters, enabling it to learn the patterns of the corresponding input sequences and output K sub-prediction results. Finally, the sub-prediction results output by the K sub-prediction models are superimposed, and the prediction information of all components is integrated to obtain the prediction result of the first load of the power grid, i.e., the first prediction result.

[0066] Specifically, the first load-related data is a time-series input sequence. ,in For a moment t The input feature vector ( D (The feature dimension includes historical load values, meteorological data, etc.) T The time length of the input sequence represents the number of input sequences. T The input data at each time point; the first prediction result is the load forecast sequence for future time points. ,in, Indicates the future number T + k The load forecast value at time , where k =1, 2, ..., H The first predictive model is a multi-layer bidirectional long short-term memory network, denoted as... ,satisfy Furthermore, for any given time... t Bidirectional LSTM (Long Short-Term Memory) networks learn features simultaneously from both forward (past to present) and backward (future to present) time sequences, and concatenate the forward and backward hidden states to obtain the time sequence. t A complete representation of its features. Furthermore, through... L Layers of bidirectional LSTM are stacked to achieve the mapping from raw features to higher-order abstract features, such as the output feature sequence of the first layer. ,in, This indicates that the first layer of bidirectional LSTM is at time [time].t The complete feature representation, This indicates the forward timing sequence in the first layer of the bidirectional LSTM at time [time value missing]. t The hidden state, This indicates the reverse timing sequence in the first layer of the bidirectional LSTM at time [time]. t The hidden state; the first l Layer (2≤ l ≤ L The input is the output feature sequence of the previous layer. The output is (Each layer contains independent bidirectional LSTM parameters), where, Indicates the first l Layer bidirectional LSTM at time t The complete feature representation is composed of forward and backward hidden states. Indicates the first l In a bidirectional LSTM layer, the forward timing is at time... t The hidden state, Indicates the first l In a layered bidirectional LSTM, the reverse timing sequence is at time... t The hidden state; the final layer outputs a high-order feature sequence as ( d L For the first L (Layer hidden dimension), where L represents the number of layers in the bidirectional LSTM; based on the features of the final layer The mapping from the fully connected layer to the prediction result can be represented as follows: ,in, To predict weights, This serves as a bias; furthermore, the model parameters are optimized by minimizing the prediction error loss function. ,in, Let the model's prediction error loss function be denoted as . The true load value at time T+k is given, and the mean square error (MSE) is used as the loss.

[0067] S203. Based on the second prediction model, the first load correlation data and the first prediction result are processed to obtain the second prediction result corresponding to the power grid.

[0068] The second prediction model is a physical information neural network model that embeds the physical constraints of power grid operation as a regularization term into the model loss function. The physical constraints are dynamically changing, ensuring that the prediction results conform to the data trend and do not violate the objective physical rules of power grid operation.

[0069] The physical information neural network model is a model that integrates data-driven approaches with physical laws. Based on traditional neural networks, it incorporates the physical laws of power grid operation (such as power conservation and line capacity limitations) to avoid unrealistic predictions, such as predicting that the line transmission power exceeds the actual maximum current carrying capacity.

[0070] The physical constraints of power grid operation are the objective physical rules that the power grid must follow, and they change dynamically with the state of the power grid. Common constraints include power conservation (node ​​input power = output power), line transmission capacity limit (power cannot exceed the maximum current carrying capacity of the line), and node voltage stability (voltage must be within the upper and lower limits).

[0071] The regularization term is a constraint clause in the model loss function. Its function is to correct the direction of model prediction, prevent the model from pursuing data fit and outputting results that violate physical laws, and ensure that the predicted value is within the range of actual operation of the power grid.

[0072] The model loss function is an indicator of whether the model's predictions are accurate. The loss function of the second prediction model consists of two parts: the error between the prediction result and the actual data, and the error of the prediction result violating the constraints of physical laws. The two are combined to guide the model optimization.

[0073] The second prediction result is the final output of the second prediction model. It is a prediction result that integrates data trends and physical laws, and it is more in line with the actual operation scenario of the power grid than the first prediction result, thus having higher reliability.

[0074] The methods for updating physical law constraints can include:

[0075] Real-time acquisition of power grid change data; updating of constraint parameters through power flow calculation based on the change data; and updating of physical law constraints based on the updated constraint parameters.

[0076] Among them, change data refers to various dynamic information that leads to changes in the physical constraints of the power grid. This may include, but is not limited to, topology change data (such as changes in the connection relationship of power grid lines and nodes), equipment parameter change data (such as changes in the performance parameters of power grid equipment itself), and operation status monitoring data (real-time operation indicators of the power grid, such as real-time line current, node voltage, equipment temperature, etc.). These data are the signal sources that trigger the update of physical constraints.

[0077] Power flow calculation is an algorithm used to analyze the operating status of a power grid. By inputting data such as power grid topology, equipment parameters, and loads, it calculates key parameters such as node voltage, line power, and losses, and serves as a calculation tool for updating constraint parameters.

[0078] Constraint parameters are specific values ​​that constitute physical law constraints. They change dynamically with the state of the power grid and may include, but are not limited to, node injected power parameters in power conservation constraints, maximum current carrying capacity parameters in line transmission capacity constraints, and upper and lower voltage limits in node voltage stability constraints.

[0079] Optionally, the physical constraints can be dynamically updated through the power grid flow calculation model, and the updated physical constraints can be re-embedded into the loss function of the second prediction model to achieve adaptive tracking of the dynamic characteristics of the power grid.

[0080] For example, the first load associated data And the first prediction result As input to the second prediction model, the first prediction result is used. The second prediction result was obtained after correction. ; where the loss function of the second prediction model This is a bi-objective optimization problem, which can be represented as:

[0081]

[0082] in, ( ) is the data loyalty loss function, ensuring that the prediction does not deviate too far from the data trend; ( ) is the physical regularity loss function, which measures the degree to which the prediction results violate physical laws; It is the dynamic regularization strength coefficient, which is an important factor in realizing dynamically changing constraints; These are neural network parameters.

[0083] Data loyalty loss function ( ) To ensure that the second prediction result does not completely discard the valuable information learned from the data by the first prediction model in order to satisfy the physical rules, the Huber loss is used to make it more robust to large errors in the first prediction result and to prevent the physical term from overcorrecting prediction points that are actually correct but appear abnormal. This can be expressed as:

[0084]

[0085] in, This indicates that the second prediction model is at time [time]. t The predicted value, This indicates that the first prediction model is at time [time]. t The predicted value, Huber loss function ( ) can be defined as:

[0086]

[0087] in, δThese are the hyperparameters of the Huber loss function; It is the prediction error ( ).

[0088] The dynamic physical residual function R() can be expressed as:

[0089]

[0090] in, The physical regularity loss function measures the degree to which the second prediction result violates the physical laws of the power grid, and is a time-related loss function. t and model parameters θ The function; Dynamic physical residual function, measuring the second prediction result With the physical state of the power grid At any moment t The degree to which physical laws are violated; It is a moment t The physical status data of the power grid includes information such as topology, equipment parameters, and operating status.

[0091] The residual function R can be composed of multiple dynamically changing sub-constraints, such as dynamic power balance residual, dynamic line power flow safety residual, and dynamic generator ramping residual; among them, dynamic power balance residual... Considering the time-varying predicted renewable energy injection and network loss estimates, it can be expressed as:

[0092]

[0093] in, This represents a row vector of all 1s, used for summing the predicted sequence or the injected power sequence; Indicates time t Predicted power injection sequence from renewable energy sources; Indicates time t Network loss sequence of the power grid.

[0094] Dynamic line power flow safety residual Medium line capacity limit It is no longer a constant, but rather changes dynamically with ambient temperature, wind speed, and other factors. By replacing the non-differentiable ReLU with a continuously differentiable penalty function (such as a polynomial barrier function), the gradient flow is smoother, which is beneficial for training. This can be expressed as:

[0095]

[0096] in, This indicates the total number of lines in the power grid; Indicates the first iThe line at the time t The trend; Indicates the first i The line at the time t The dynamic capacity limit varies with factors such as ambient temperature and wind speed; This represents the hyperparameter of the penalty function, which is used to adjust the degree of penalty when the power flow of the line approaches the capacity limit.

[0097] Dynamic generator climbing residual Gradient limitations of generators It is also dynamically adjusted based on the real-time status of the unit (such as hot start, cold start). This constraint is directly integrated into the prediction of continuous time points, and can be expressed as:

[0098]

[0099] in, It is the first j The generator at time t To meet load forecasting And the allocated effort; Indicates the first j The generator at time contribution; This represents the time interval between two adjacent moments; Indicates the first j The generator at time t The dynamic ramp rate limit is adjusted according to the real-time status of the unit (such as hot start and cold start).

[0100] To make the physical constraints truly change dynamically, let λ itself be a function determined by a lightweight network or rules, which can be expressed as:

[0101]

[0102] in, It is the basic regularization strength; This is the sensitivity coefficient; Used to evaluate the first prediction result At any moment t Functions of uncertainty, such as the standard deviation of an ensemble model and the width of the predicted probability interval.

[0103] For example, real-time data on changes in the power grid can be acquired through sensors and monitoring systems, including whether the grid topology has changed (e.g., a line is disconnected due to maintenance), whether equipment parameters have been adjusted (e.g., transformer capacity is upgraded), and whether the operating status is abnormal (e.g., line current is close to its upper limit). Furthermore, based on the collected change data, the power flow calculation model is invoked to recalculate the power grid operating status. For instance, if a line's maximum current carrying capacity decreases due to increased temperature (change data), a new maximum current carrying capacity parameter can be calculated through power flow calculation. If a new distributed power source is added to a node (topology / equipment change data), the injected power parameter of that node can be updated through power flow calculation. This translates grid changes into specific adjustments to constraint parameters. Subsequently, the updated constraint parameters can be used to redefine the specific content of physical constraint rules. For example, line transmission capacity constraints can be updated from ≤100MW (old parameter) to ≤80MW (new parameter), and node voltage constraints can be updated from 0.95-1.05kV (old parameter) to 0.92-1.08kV (new parameter), ultimately ensuring that physical constraint rules perfectly match the current state of the power grid.

[0104] It should be noted that the training process for the first and second prediction models includes:

[0105] Based on time series cross-validation, historical load-related data are divided into training, validation, and test sets. An optimizer is used to train the initial first prediction model, and physical law constraint loss terms are introduced in real time during the training of the initial second prediction model. If the initial first prediction model achieves the first prediction accuracy index, the second prediction model is obtained.

[0106] Optionally, the optimizer can be the Adam optimizer; the learning rates in the initial first prediction model and the initial second prediction model can be dynamically adjusted using a cosine annealing strategy; gradient pruning is used to prevent gradient explosion during model training; and an early stopping strategy can also be used to determine the optimal training epochs for the model.

[0107] Among them, time series cross-validation is a more rigorous data partitioning method for time series data; the training set is the data on which the model learns, and the model directly adjusts its internal parameters (weights and biases) on this data to learn the patterns; the validation set is the data used for testing, that is, during the training process, this part of the data is used to evaluate the model performance, to adjust hyperparameters (such as learning rate, number of network layers) and to decide when to stop training (early stop), the model does not learn directly from the validation set; the test set is the data used to evaluate the model. After the model training and tuning are completely finished, the test set is used to objectively and fairly evaluate the model's final generalization ability.

[0108] An optimizer guides how model parameters are updated by calculating gradients (which indicate the direction and magnitude of parameter adjustments) to minimize the value of the loss function. It acts as navigation software or a coach during the model training process.

[0109] Cosine annealing is a learning schedule that dynamically adjusts the learning rate. The process is that as the number of training rounds increases, the learning rate follows the shape of a cosine function, first slowly decreasing from the initial value, then rapidly decreasing in the middle, and finally slowly approaching 0 at the end.

[0110] Gradient clipping acts as a safety valve during training. When the gradient (adjustment step size) of the model parameters is too large, it limits it to a reasonable threshold to prevent the model parameters from oscillating violently or even exploding (numerical overflow) due to excessively large single update step size, thereby ensuring the stability of the training process.

[0111] Early stopping is a mechanism to prevent the model from overlearning or rote memorization. The process involves continuously monitoring the model's performance on the validation set during training. Once the performance on the validation set stops improving or even starts to decline after several rounds of training, training is terminated early. This can prevent the model from overfitting the training data, thereby obtaining a model with stronger generalization ability and saving training time.

[0112] The first prediction accuracy metric is a pre-set performance threshold used to determine whether the first prediction model is qualified; the second prediction accuracy metric is a pre-set performance threshold used to determine whether the first prediction model is qualified.

[0113] S204. Based on the first and second prediction results, the third prediction result corresponding to the power grid is obtained.

[0114] One possible approach is to concatenate the first and second prediction results by time steps to form a fused feature vector; input the fused feature vector into an attention network to obtain the dynamic weights of the first and second prediction results at each time step; and based on the dynamic weights, perform a weighted summation of the first and second prediction results to obtain the third prediction result corresponding to the power grid.

[0115] Among them, the third prediction result is a consensus prediction or elite prediction formed by combining the advantages of the first prediction result (purely data-driven) and the second prediction result (physical information constraint), which is a more accurate and robust final result; the dimension of the fused feature vector is the sum of the feature dimensions of the first prediction result and the second prediction result; the attention network is a special neural network mechanism that can automatically evaluate the importance of different parts of the input information and assign them appropriate attention weights; the weighted summation is to multiply each data point by its corresponding weight and then add them together to obtain the final result.

[0116] For example, the first prediction result Second prediction results At each identical time step t, the data is concatenated to form a fused feature vector. , Furthermore, this fused feature vector [ , The input is fed into an attention network, which performs a comprehensive analysis based on the information contained in the current fused feature vector and outputs a dynamic weight for this time step. This weight represents the confidence level of the first prediction at that moment; furthermore, at each time step t, the calculated dynamic weight is used... Regarding the first prediction result Second prediction results A weighted sum is performed to obtain the final third prediction result. The calculation formula can be: .

[0117] Furthermore, after obtaining the third prediction result, it can be processed, and the processing steps may include:

[0118] The third prediction result is smoothed using a moving average algorithm based on an adaptive window size; the smoothed third prediction result is then corrected based on the safety boundary conditions of the power grid operation.

[0119] The adaptive window size can be dynamically adjusted based on the proportion of high-frequency noise energy in the predicted sequence. For example, the power spectral density of the predicted result within a preset time band is calculated. When the proportion of high-frequency energy (such as fluctuation components within 1 hour) exceeds a threshold, the window size is increased to enhance the smoothing effect; when the proportion of high-frequency energy is lower than the threshold, the window size is decreased to retain effective details. Through this sliding window smoothing process, high-frequency disturbances caused by model fitting errors or data noise in the predicted result are eliminated, while maintaining the trend characteristics and key abrupt change points of the load sequence. The safety boundary conditions may include, but are not limited to, the maximum load limit and the minimum load threshold.

[0120] For example, the proportion of high-frequency energy can be determined by calculating the power spectral density of the prediction results within a preset time frequency band (such as the fluctuation component within 1 hour). For instance, if the proportion of high-frequency energy exceeds a threshold, it indicates that the high-frequency disturbances caused by noise or model fitting errors are strong. In this case, the window size is increased to allow more data to participate in averaging, enhance the smoothing effect, and eliminate these disturbances. If the proportion of high-frequency energy is lower than the threshold, it indicates that the high-frequency fluctuations are mostly effective details (such as key abrupt changes in load). In this case, the window size is reduced to avoid over-smoothing and retain these effective information. Finally, this eliminates the high-frequency disturbances in the prediction results caused by model fitting errors or data noise, while maintaining the trend characteristics of the load sequence (such as the overall load change trend) and key abrupt changes (such as nodes where the load suddenly rises or falls sharply).

[0121] Safety boundary conditions include, but are not limited to, maximum load limits (the maximum load that grid equipment can bear; exceeding this limit will cause overload or failure) and minimum load thresholds (too low a load may lead to low grid operating efficiency or stability problems). The smoothed prediction results can be compared with these safety boundaries. If the prediction value exceeds the maximum load limit, it is adjusted down to within the limit; if it is lower than the minimum load threshold, it is adjusted up to above the threshold to ensure that the prediction results are within a reasonable range for safe grid operation.

[0122] The aforementioned intelligent power grid load forecasting method based on multi-model collaboration obtains the first load-related data corresponding to the power grid, providing a data foundation for subsequent power grid forecasting. Furthermore, based on a first forecasting model constructed from a multi-layer bidirectional long short-term memory network, the first load-related data is processed to obtain the first forecasting result corresponding to the power grid, achieving preliminary power grid forecasting. Further, based on a second forecasting model that embeds dynamically changing physical constraints as regularization terms into the model loss function, the first load-related data and the first forecasting result are processed to obtain the second forecasting result corresponding to the power grid, achieving adjustment of the first forecasting result. Subsequently, a third forecasting result corresponding to the power grid can be obtained based on the first and second forecasting results. This scheme introduces a first forecasting model to perform preliminary load forecasting for the power grid; furthermore, by introducing a second forecasting model that embeds dynamically changing physical constraints as regularization terms into the model loss function, it considers the dynamic changes of the power grid, laying the foundation for accurate load forecasting; finally, by combining the first and second forecasting results, an accurate load forecasting result is obtained.

[0123] The above text combined Figures 1 to 2 The present application provides a detailed description of the intelligent power grid load forecasting method based on multi-model collaboration. The apparatus and equipment provided in the present application will be described below with reference to the accompanying drawings.

[0124] like Figure 3 As shown in the figure, this is a structural block diagram of a multi-model collaborative intelligent power grid load forecasting device 600 provided in an embodiment of this application. The multi-model collaborative intelligent power grid load forecasting device 600 includes: an acquisition module 601, a first forecasting module 602, a second forecasting module 603, and a third forecasting module 604, wherein:

[0125] The acquisition module 601 is used to acquire the first load-related data corresponding to the power grid;

[0126] The first prediction module 602 is used to process the first load-related data based on the first prediction model to obtain the first prediction result corresponding to the power grid; wherein, the first prediction model is a data-driven model constructed based on a multi-layer bidirectional long short-term memory network.

[0127] The second prediction module 603 is used to process the first load-related data and the first prediction result based on the second prediction model to obtain the second prediction result corresponding to the power grid; wherein, the second prediction model is a physical information neural network model that embeds the physical law constraints of power grid operation as regularization terms into the model loss function; the physical law constraints are dynamically changing;

[0128] The third prediction module 604 is used to obtain the third prediction result corresponding to the power grid based on the first prediction result and the second prediction result.

[0129] In one embodiment, the third prediction module 604 is specifically used for:

[0130] The first and second prediction results are concatenated according to time steps to form a fused feature vector;

[0131] The fused feature vector is input into the attention network to obtain the dynamic weights of the first and second prediction results at each time step.

[0132] Based on dynamic weights, the first and second prediction results are weighted and summed to obtain the third prediction result corresponding to the power grid.

[0133] In one embodiment, the first prediction module 602 is specifically used for:

[0134] Based on the variational mode decomposition algorithm, the load sequence in the historical load data is decomposed to obtain K eigenmode function components; where K is a preset positive integer.

[0135] Based on the K intrinsic mode function components and the non-load characteristic data in the first load correlation data, the first prediction result corresponding to the power grid is obtained.

[0136] In one embodiment, the first prediction module 602 is specifically used for:

[0137] Trend terms are extracted from the K intrinsic mode function components, and the effective components are retained after removing noise components.

[0138] The effective components are aligned with the non-load characteristic data in the first load-related data according to time steps and then concatenated to form K new input sequences.

[0139] K new input sequences are fed into K parallel sub-prediction models, each of which is a bidirectional long short-term memory network with independently optimized structural parameters in the first prediction model.

[0140] The K sub-prediction results are superimposed to obtain the first prediction result.

[0141] In one embodiment, the smart grid load forecasting device 600 based on multi-model collaboration further includes:

[0142] The training module is used to divide historical load association data into training, validation and test sets based on time series cross-validation; to train the initial first prediction model using an optimizer, and to introduce physical law constraint loss terms in real time during the training of the initial second prediction model; to obtain the second prediction model when the initial first prediction model reaches the first prediction accuracy index, and when the initial second prediction model reaches the second prediction accuracy index.

[0143] In one embodiment, the smart grid load forecasting device 600 based on multi-model collaboration further includes:

[0144] The constraint update module is used to collect real-time changes in power grid data; based on the changes in data, it updates the constraint parameters through power flow calculation; and based on the updated constraint parameters, it updates the physical law constraints.

[0145] In one embodiment, the smart grid load forecasting device 600 based on multi-model collaboration further includes:

[0146] The processing module is used to smooth the third prediction result based on the moving average algorithm with an adaptive window size; and to correct the smoothed third prediction result based on the safety boundary conditions of the power grid operation.

[0147] The intelligent power grid load forecasting device 600 based on multi-model collaboration according to the embodiments of this application can correspond to the execution of the method described in the embodiments of this application, and the other operations and / or functions of each module / unit of the intelligent power grid load forecasting device 600 based on multi-model collaboration are respectively for implementing Figure 2 For the sake of brevity, the corresponding processes of each method in the illustrated embodiments will not be described in detail here.

[0148] This application also provides a computing device. This computing device can be a local computing device or an application server.

[0149] like Figure 4 As shown in the figure, this is an internal structural diagram of a computing device provided in an embodiment of this application. The computing device 700 includes a bus 701, a processor 702, a communication interface 703, and a memory 704. The processor 702, the memory 704, and the communication interface 703 communicate with each other via the bus 701.

[0150] The 701 bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0151] The processor 702 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0152] Communication interface 703 is used for external communication. For example, communication interface 703 can be used to communicate with terminal 102. Communication interface 703 is used to send a third prediction result to terminal 102 so that terminal 102 can display the third prediction result.

[0153] Memory 704 may include volatile memory, such as random access memory (RAM). Memory 704 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0154] The memory 704 stores executable code, and the processor 702 executes the executable code to perform the aforementioned intelligent power grid load forecasting method based on multi-model collaboration.

[0155] Specifically, in achieving Figure 3 In the case of the illustrated embodiment, and Figure 3 When the modules or units of the intelligent power grid load forecasting device based on multi-model collaboration described in the embodiments are implemented in software, the execution... Figure 3 The software or program code required for the functions of each module / unit can be partially or entirely stored in the memory 704. The processor 702 executes the program code corresponding to each unit stored in the memory 704, and executes the aforementioned intelligent power grid load forecasting method based on multi-model collaboration.

[0156] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium capable of being stored by a computing device, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to execute the aforementioned intelligent power grid load forecasting method based on multi-model collaboration.

[0157] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0158] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0159] When the computer program product is executed by a computer, the computer executes any of the aforementioned methods of the intelligent power grid load forecasting method based on multi-model collaboration. The computer program product can be a software installation package; when any of the aforementioned methods of the intelligent power grid load forecasting method based on multi-model collaboration are required, the computer program product can be downloaded and executed on the computer.

[0160] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0161] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of protection of this application.

Claims

1. A method for intelligent power grid load forecasting based on multi-model collaboration, characterized in that, The method includes: Obtain the first load-related data corresponding to the power grid; wherein, the first load-related data includes historical load data; Based on the variational mode decomposition algorithm, the load sequence in the historical load data is decomposed to obtain K eigenmode function components; where K is a preset positive integer. Trend terms are extracted from the K intrinsic mode function components, and the effective components are retained after removing noise components. The effective components are aligned with the non-load feature data in the first load-related data according to time steps and then concatenated to form K new input sequences. K new input sequences are input into K parallel sub-prediction models to obtain K sub-prediction results; wherein, the sub-prediction model is a bidirectional long short-term memory network with independently optimized structural parameters in the first prediction model; the first prediction model is a data-driven model built on a multi-layer bidirectional long short-term memory network. The K sub-prediction results are superimposed to obtain the first prediction result; Based on the second prediction model, the first load-related data and the first prediction result are processed to obtain the second prediction result corresponding to the power grid; wherein, the second prediction model is a physical information neural network model that embeds the physical law constraints of power grid operation as a regularization term into the model loss function; the physical law constraints are dynamically changing; Based on the first and second prediction results, a third prediction result corresponding to the power grid is obtained; The loss function of the second prediction model It is a bi-objective optimization problem: in, ( ) is the data loyalty loss function; ( ) is the physical regularity loss function; It is the dynamic regularization strength coefficient; These are the neural network parameters; t is the time step. Based on the first and second prediction results, a third prediction result corresponding to the power grid is obtained, including: The first prediction result and the second prediction result are concatenated according to time steps to form a fused feature vector; The fused feature vector is input into the attention network to obtain the dynamic weights of the first prediction result and the second prediction result at each time step. Based on the dynamic weights, the first prediction result and the second prediction result are weighted and summed to obtain the third prediction result corresponding to the power grid.

2. The method according to claim 1, characterized in that, The training process for the first and second prediction models includes: Based on time series cross-validation, historical load correlation data are divided into training set, validation set and test set; The initial first prediction model is trained using an optimizer, and the initial second prediction model is trained in real time by introducing a loss term constrained by physical laws. If the initial first prediction model achieves a first prediction accuracy index, a first prediction model is obtained; and if the initial second prediction model achieves a second prediction accuracy index, a second prediction model is obtained.

3. The method according to claim 1, characterized in that, The methods for updating the physical law constraints include: Real-time acquisition of power grid change data; Based on the changed data, the constraint parameters are updated through power grid power flow calculation; The physical law constraints are updated based on the updated constraint parameters.

4. The method according to claim 1, characterized in that, The post-processing steps for the third prediction result include: The third prediction result is smoothed using a moving average algorithm based on an adaptive window size. Based on the safety boundary conditions of power grid operation, the smoothed third prediction result is corrected.

5. A smart power grid load forecasting device based on multi-model collaboration, characterized in that, The device includes: The acquisition module is used to acquire the first load-related data corresponding to the power grid; wherein, the first load-related data includes historical load data; The first prediction module is used to process the first load-related data based on the first prediction model to obtain the first prediction result corresponding to the power grid; wherein, the first prediction model is a data-driven model built on a multi-layer bidirectional long short-term memory network. The second prediction module is used to process the first load-related data and the first prediction result based on the second prediction model to obtain the second prediction result corresponding to the power grid; wherein, the second prediction model is a physical information neural network model that embeds the physical law constraints of power grid operation as a regularization term into the model loss function; the physical law constraints are dynamically changing; the loss function of the second prediction model... It is a bi-objective optimization problem: in, ( ) is the data loyalty loss function; ( ) is the physical regularity loss function; It is the dynamic regularization strength coefficient; These are the neural network parameters; t is the time step. The third prediction module is used to obtain a third prediction result corresponding to the power grid based on the first prediction result and the second prediction result; The first prediction module is used to decompose the load sequence in historical load data based on the variational mode decomposition algorithm to obtain K intrinsic mode function components; where K is a preset positive integer; extract the trend term from the K intrinsic mode function components, remove noise components and retain the effective components; align the effective components with the non-load feature data in the first load association data according to time steps and concatenate them to form K new input sequences; input the K new input sequences into K parallel sub-prediction models respectively to obtain K sub-prediction results; wherein the sub-prediction model is a bidirectional long short-term memory network with independently optimized structural parameters in the first prediction model; and superimpose the K sub-prediction results to obtain the first prediction result. The third prediction module is used to concatenate the first prediction result and the second prediction result according to time steps to form a fused feature vector; input the fused feature vector into an attention network to obtain the dynamic weights of the first prediction result and the second prediction result at each time step; and perform a weighted summation of the first prediction result and the second prediction result based on the dynamic weights to obtain the third prediction result corresponding to the power grid.

6. A computing device, characterized in that, Including memory and processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method as described in any one of claims 1 to 4.

Images