Building baseline load prediction model training method and system, and load prediction method

Abstract

This application relates to a training method, system, and load forecasting method for a building baseline load forecasting model. The method includes: acquiring a candidate feature set and a historical load training set; calling an initial model to predict the training set based on the candidate feature set, obtaining the prediction error, and extracting the gain and coverage indices of the candidate feature set to determine its importance; iteratively eliminating candidate feature sets by combining the prediction error and importance to obtain the optimal features; training a target model using the optimal features and the training set; using the target model to predict the baseline load during demand response, calculating the load reduction based on the actual operating load, and generating corresponding demand response scheduling instructions. This application achieves interactive collaboration between feature selection and the prediction model, improving the prediction accuracy of building baseline load and providing reliable data support for the stable scheduling of power grid physical equipment.

Description

Technical Field

[0001] This application relates to the field of power technology, and in particular to a training method for a building baseline load forecasting model, a training system for a building baseline load forecasting model, a load forecasting method, a load forecasting system, and computer equipment. Background Technology

[0002] With the continuous development and construction of new power systems, demand response has become an effective means of maintaining the balance between power grid supply and demand. As a major energy consumer, the accurate prediction of the baseline load of buildings is of great reference value for the smooth implementation of demand response and the stable operation of the power grid.

[0003] Currently, baseline load forecasting typically employs traditional statistical methods or various deep learning models for calculation. In the feature selection stage of data processing, an independent filtering feature selection algorithm is used to unidirectionally screen candidate features, and the selected features are then directly input into the prediction model for subsequent calculations.

[0004] This leads to a disconnect between existing feature selection methods and prediction models. The selected features are often not optimally configured for the final prediction model, resulting in low prediction accuracy. Summary of the Invention

[0005] Based on this, it is necessary to provide a building baseline load forecasting model training method, a building baseline load forecasting model training system, a load forecasting method, a load forecasting system, and computer equipment that can achieve adaptive optimization of feature selection and baseline load forecasting to improve forecasting accuracy, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides a method for training a building baseline load prediction model, including:

[0007] Obtain the candidate feature set and historical load training set of the target building;

[0008] The initial model is invoked to predict the historical load training set based on the candidate feature set to obtain the prediction error. The gain index and coverage index of each feature in the candidate feature set are extracted. Based on the gain index and the coverage index, the importance index of each feature is determined.

[0009] By combining the prediction error and the importance index, the candidate feature set is iteratively eliminated to obtain the optimal feature subset;

[0010] The optimal feature subset and the historical load training set are input into the initial model for training to obtain the target model; the target model is used to receive online feature data of the target building, extract target data corresponding to the optimal feature subset from the online feature data for prediction, and obtain the baseline load prediction result.

[0011] In one embodiment, obtaining the candidate feature set and historical load training set of the target building includes:

[0012] Obtain the initial load data of the target building during the historical demand response process, as well as the time status and user participation status corresponding to the initial load data;

[0013] Based on the time status and the user participation status, the initial load data is classified to obtain multiple load data subsets;

[0014] Extract the load data subset that is in the target state from the multiple load data subsets, and use it as the historical load training set.

[0015] In one embodiment, the time state includes an unknown response phase, a known response phase, and an execution response phase;

[0016] The user participation status includes no bidding status, failed bidding status, and successful bidding status;

[0017] The step of extracting the load data subset in the target state from the multiple load data subsets as the historical load training set includes:

[0018] The subset of load data that is in the unknown response stage and corresponds to the non-bid, bid failure, or bid success state will be used as the historical load training set.

[0019] In one embodiment, the initial model is an ensemble tree model;

[0020] The extraction of gain and coverage metrics for each feature in the candidate feature set includes:

[0021] During the prediction process, the first and second derivative data of the prediction loss function are extracted;

[0022] Based on the second derivative data, the influence of node samples when the feature is split is calculated and used as the coverage index.

[0023] Based on the first-order derivative data and the second-order derivative data, the loss reduction caused by the feature split is calculated and used as the gain index.

[0024] In one embodiment, the step of iteratively eliminating candidate feature sets by combining the prediction error and the importance index to obtain an optimal feature subset includes:

[0025] Sort the features in the candidate feature set in ascending order according to the importance index;

[0026] The features at the end of the sort are removed sequentially, and the prediction error re-output by the initial model after removing the features is obtained.

[0027] If the prediction error is detected to no longer decrease or the number of remaining features is less than a preset constraint value, the elimination process stops, and the remaining features are determined as the optimal feature subset.

[0028] Secondly, this application also provides a training system for a building baseline load prediction model, comprising:

[0029] The acquisition module is used to acquire the candidate feature set and historical load training set of the target building;

[0030] The processing module is used to call the initial model to predict the historical load training set based on the candidate feature set, obtain the prediction error, extract the gain index and coverage index of each feature in the candidate feature set, and determine the importance index of each feature based on the gain index and the coverage index.

[0031] The elimination module is used to combine the prediction error and the importance index to iteratively eliminate the candidate feature set to obtain the optimal feature subset;

[0032] The training module is used to input the optimal feature subset and the historical load training set into the initial model for training to obtain the target model; the target model is used to receive the online feature data of the target building, extract the target data corresponding to the optimal feature subset from the online feature data for prediction, and obtain the baseline load prediction result.

[0033] Thirdly, this application provides a load forecasting method, including:

[0034] Acquire online characteristic data of the target building during demand response;

[0035] Extract the target data corresponding to the optimal feature from the online feature data;

[0036] The target data is input into a pre-trained target model for prediction to obtain the baseline load prediction result of the target building;

[0037] Obtain the actual operating load of the target building during the demand response period;

[0038] Based on the baseline load forecast results and the actual operating load, the load reduction amount for the target building is determined, and a corresponding demand response scheduling instruction is generated according to the load reduction amount.

[0039] The optimal features and the target model are obtained through the training method of the baseline load prediction model as described in any of the above.

[0040] In one embodiment, generating a corresponding demand response scheduling instruction based on the load reduction amount includes:

[0041] Obtain the target reduction threshold for the target building during the demand response period;

[0042] If it is determined that the load reduction amount is less than the target reduction threshold, a demand response scheduling instruction is generated based on the difference between the target reduction threshold and the load reduction amount to reduce the operating power of the target energy-consuming equipment.

[0043] If the load reduction amount is determined to be greater than or equal to the target reduction threshold, a scheduling instruction is generated to maintain the current operating state of the target energy-consuming equipment.

[0044] Fourthly, this application provides a load forecasting system, comprising:

[0045] The feature acquisition module is used to acquire online feature data of the target building during the demand response period;

[0046] The feature extraction module is used to extract the target data corresponding to the optimal feature from the online feature data;

[0047] The load forecasting module is used to input the target data into a pre-trained target model for forecasting, and obtain the baseline load forecasting result of the target building;

[0048] The load acquisition module is used to acquire the actual operating load of the target building during the demand response period;

[0049] The scheduling generation module is used to determine the load reduction amount of the target building based on the baseline load forecast results and the actual operating load, and generate corresponding demand response scheduling instructions according to the load reduction amount.

[0050] The optimal features and the target model are obtained through the training method of the building baseline load prediction model as described in any of the above.

[0051] Fifthly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method. Specifically, it implements the steps of the building baseline load forecasting model training method described in any of the above claims, or the steps of the load forecasting method described in any of the above claims.

[0052] The aforementioned building baseline load forecasting model training method, building baseline load forecasting model training system, load forecasting method, load forecasting system, and computer equipment achieve adaptive interactive collaboration between feature selection and the underlying logic of the forecasting model by utilizing the initial model's own calculation indicators in conjunction with prediction errors for closed-loop iterative elimination. This overcomes the shortcomings of traditional filtering-based feature selection being disconnected from the forecasting model, avoids the omission of important features or the introduction of irrelevant features, and improves the accuracy of baseline load forecasting. At the same time, by combining the forecasting model with the generation of demand response scheduling instructions, closed-loop control from simple data forecasting to physical power grid equipment scheduling is achieved, providing reliable data support for the power consumption strategy of the power grid dispatching center and improving the stability of the new power system operation. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a diagram illustrating the application environment of power demand response control in one embodiment.

[0055] Figure 2 This is a schematic diagram of the training process for a building baseline load prediction model in one embodiment;

[0056] Figure 3 This is a time-axis diagram of power demand response in one embodiment;

[0057] Figure 4 This is a structural block diagram of a building baseline load prediction model system in one embodiment;

[0058] Figure 5 This is a flowchart illustrating a load forecasting method in one embodiment;

[0059] Figure 6 This is a timing diagram illustrating building baseline load prediction training and predictive control in one embodiment;

[0060] Figure 7Here is a block diagram of a load forecasting system in one embodiment;

[0061] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0064] The load forecasting method provided in this application can be applied to, for example, Figure 1 The power demand response control application environment shown includes a data acquisition device 110, a data processing device 120, and an energy consumption control device 130.

[0065] The data acquisition device 110 is connected to the data processing device 120 via a communication network. The data acquisition device 110 may include a smart meter or a temperature and humidity sensor, and the communication network may include a wired fiber optic network or a wireless mobile network. The data acquisition device 110 collects online characteristic data and historical load data of the target building. The online characteristic data may include ambient temperature data or equipment operating load data, and sends the data to the data processing device 120.

[0066] The data processing device 120 is connected to the energy consumption control device 130 via a communication network. The data processing device 120 may include a cloud data center or an edge computing gateway. The data processing device 120 receives data sent by the data acquisition device 110, is responsible for performing feature selection, model training, and baseline load forecasting, and generates demand response scheduling instructions, which are then sent to the energy consumption control device 130.

[0067] The energy consumption control device 130 is connected to the target energy-consuming device. The energy consumption control device 130 may include a central air conditioning controller or a lighting concentrator, and the target energy-consuming device may include a central air conditioning unit or a building lighting circuit. The energy consumption control device 130 receives demand response scheduling instructions issued by the data processing device 120 and adjusts the operating status of the target energy-consuming device accordingly to achieve physical load reduction.

[0068] In one exemplary embodiment, such as Figure 2 As shown, a method for training a building baseline load prediction model is provided, which can be applied to... Figure 1 Taking the data processing device 120 as an example, the explanation includes the following steps 210 to 240. Wherein:

[0069] Step 210: Obtain the candidate feature set and historical load training set of the target building.

[0070] The candidate feature set can be a collection of external meteorological factors or internal operating parameters that affect the building's electricity load, such as outdoor temperature parameters, solar radiation intensity, air conditioning load parameters, or lighting socket load parameters. The historical load training set can be the actual electricity load records of the target building over a past period, such as electricity load data during historical unknown response phases or electricity load data during historical bidding failure states.

[0071] For example, the pre-stored feature data and load records of the target building are read from a local database or external data source through a data interface and loaded into memory as a candidate feature set and a historical load training set.

[0072] Step 220: Call the initial model to predict the historical load training set based on the candidate feature set, obtain the prediction error, and extract the gain index and coverage index of each feature in the candidate feature set. Based on the gain index and coverage index, determine the importance index of each feature.

[0073] The initial model can be a machine learning algorithm model capable of calculating feature metrics, such as an ensemble tree model or a decision tree model. Prediction error can be the degree of difference between the model's predicted values ​​and the historical true load values. Gain metric can be the amount of loss function reduction brought about by a feature splitting the model. Coverage metric can be the influence of node samples when a feature is split. Importance metric can be a comprehensive quantitative score measuring the degree of influence of each candidate feature on the prediction result.

[0074] For example, the candidate feature set is input into the initial model, and the load prediction value corresponding to each candidate feature is output. The load prediction value is compared with the true value in the historical load training set to calculate the prediction error corresponding to each candidate feature. Simultaneously, during the internal computation of the initial model, the first and second derivative data of the model's loss function are extracted. The coverage index is calculated using the second derivative data, and the gain index is calculated using the first and second derivative data. The obtained gain and coverage indices are normalized to unify them to the same numerical magnitude. Subsequently, according to the pre-configured index weights, a weighted average is calculated on the normalized gain and coverage indices to obtain the final importance score corresponding to each candidate feature.

[0075] Step 230: Combine prediction error and importance index to iteratively eliminate candidate feature sets and obtain the optimal feature subset.

[0076] The optimal feature subset can be the core feature set that has the greatest positive effect on baseline load prediction after screening, such as retained temperature features or historical load features.

[0077] For example, candidate features are sorted in ascending order based on the calculated importance index. Following the sorting, the least important features are sequentially removed from the candidate feature set. After each removal operation, the initial model is triggered to re-predict and output the updated prediction error. The removal process continues until the prediction error after removing features no longer decreases, at which point the removal operation stops, and the remaining features in the set at the time of stopping are determined as the optimal features. In some embodiments, if the number of remaining features is detected to be less than a pre-configured constraint value, the removal operation also stops, and the optimal features are determined.

[0078] Step 240: Input the optimal feature subset and historical load training set into the initial model for training to obtain the target model. The target model is used to receive online feature data of the target building, extract target data corresponding to the optimal feature subset from the online feature data for prediction, and obtain the baseline load prediction result. Based on the baseline load prediction result and the actual operating load, determine the load reduction amount and generate the corresponding demand response scheduling instruction.

[0079] The target model can be the final model with updated parameters that can be directly used for online forecasting. The actual operating load can be the actual power consumption of the target building during demand response. The load reduction can be the difference between the baseline load forecast and the actual operating load. Demand response scheduling instructions can be machine control messages used to regulate the operating status of energy-consuming equipment, such as temperature regulation instructions or power limiting instructions.

[0080] For example, the data dimension of the optimal feature corresponding to the optimal feature subset in the historical load training set is extracted and input into the initial model for multiple rounds of parameter iteration updates. When the training termination condition is met, the model parameters are fixed to generate the target model. In the online application phase, online feature data of the target building during demand response is acquired, and the target data corresponding to the optimal feature is extracted and input into the target model to output the baseline load prediction result. Subsequently, the actual operating load is acquired, and the difference between the baseline load prediction result and the actual operating load is calculated to obtain the load reduction amount. The load reduction amount is compared with the pre-configured target reduction threshold to generate the corresponding demand response scheduling instruction and send it to the underlying energy consumption control device 130. In some embodiments, if it is determined that the load reduction amount is greater than or equal to the target reduction threshold, a scheduling instruction to maintain the current operating state is generated.

[0081] In the above-mentioned training method for the building baseline load forecasting model, by acquiring candidate feature sets and historical load training sets, calling the initial model for prediction and extracting gain and coverage indices, then calculating importance indices, and combining prediction error iterative elimination to obtain the optimal feature subset, the target model is finally trained and demand response scheduling instructions are generated. This achieves adaptive and coordinated adjustment between feature selection and the underlying logic of the prediction model, overcomes the defect of traditional filtering feature selection being disconnected from the prediction model, and improves the accuracy of baseline load forecasting. At the same time, it realizes closed-loop control from simple data prediction to physical power grid equipment scheduling, providing reliable data support for the power consumption strategy of the power grid dispatching center and improving the stability of the new power system operation.

[0082] In an exemplary embodiment, step 210 involves obtaining the candidate feature set and historical load training set of the target building, including steps 212 to 216. Wherein:

[0083] Step 212: Obtain the initial load data of the target building during the historical demand response process, as well as the time status and user participation status corresponding to the initial load data.

[0084] The initial load data can be the original power consumption records of the target building when it participated in grid dispatch in the past, such as smart meter measurement data or building master meter billing data. The time status can be the specific service stage of the demand response event on the timeline, such as the unknown response stage or the execution response stage. The user participation status can be the target building owner's intention and outcome in the market bidding process, such as the non-bidding status or the bidding status.

[0085] For example, historical electricity load records for the target building within a set time period are retrieved from a local database or an external electricity marketing platform via a communication network. Simultaneously, historical demand response event logs issued by the power grid dispatch center are obtained. The historical electricity load records and historical demand response event logs are timestamped, and the time status label and user participation status label corresponding to each record are extracted.

[0086] Step 214: Classify the initial load data according to the time status and user participation status to obtain multiple load data subsets.

[0087] The load data subset can be a set of load data with the same time status and user participation status labels, such as a set of load data with unknown response and no bidding, or a set of load data with executed response and successful bidding.

[0088] For example, each piece of initial load data is traversed, and the accompanying time status label and user participation status label are read. A cross-classification logical matrix is ​​constructed based on the combination of these two label dimensions. Initial load data with the same label combination are grouped into the same matrix cell, forming multiple independent and non-overlapping subsets of load data.

[0089] Step 216: Extract the load data subsets that are in the target state from multiple load data subsets and use them as the historical load training set.

[0090] The target state can be a specific combination of states that reflects the actual electricity consumption habits of the target building and is not subject to human manipulation. The historical load training set can be the basic data samples used for subsequent parameter optimization in the input model.

[0091] For example, in the cross-classification logic matrix, the matrix cell whose label combination corresponds to the target state is located. The entire subset of load data contained in this cell is copied and extracted, and stored in the model training database as a historical load training set specifically used for training the baseline load prediction model. In some embodiments, if it is determined that the amount of data in the extracted load data subset is less than a pre-configured training requirement threshold, a data resampling action is triggered, or load data from an earlier historical time period is obtained for cross-classification to supplement the historical load training set.

[0092] In this embodiment, by acquiring initial load data and its corresponding time status and user participation status, cross-classifying based on status labels and extracting a subset of load data under the target status as a historical load training set, the system effectively isolates and removes data interfered with by deliberate manipulation, thereby ensuring the objectivity and reliability of the model training data and avoiding load prediction distortion caused by historical basic data pollution.

[0093] In one embodiment, such as Figure 3 The timeline of electricity demand response shows that the time states include the unknown response stage, the known response stage, and the execution response stage; the user participation states include the non-bid state, the bid failure state, and the bid success state.

[0094] Specifically, during the unknown response phase, users have not yet received any notification that a demand response is required, and their electricity consumption behavior is in a completely natural state without external interference. The historical load data during this period is real and has not been deliberately adjusted, making it suitable for calculating the user's baseline load.

[0095] During the known response phase, with the release of the demand response notification, users are informed that subsidies will be available at a certain time in the future. At this point, some users, in order to earn more subsidies, deliberately artificially inflate their electricity load during this period to raise their baseline data, thus making it appear that they have reduced their load more during the actual implementation. Furthermore, as the process progresses, the user group begins to stratify: users in the non-bidding state (those who received the notification but did not participate in the application) and users in the bidding failure state (those who participated in the market application and bidding but failed to win the bid due to excessively high prices or other reasons).

[0096] During the response execution phase, within the time window specified by the power grid, the winning bidders formally implement peak shaving or valley filling load adjustments. After execution, the actual load data during this period will be compared with the baseline load calculated in the first phase to conduct post-event evaluation and financial settlement. At this stage, the user group consists of those who successfully bid.

[0097] When constructing a forecasting model or load forecasting model, data from known response phases should be removed or outliers cleaned and not used as training sets or historical reference standards to ensure the fairness of the load forecasting and evaluation system and avoid the influence of data distortion.

[0098] Step 216 extracts load data subsets from multiple load data subsets that are in the target state, as the historical load training set. This includes: using load data subsets that are in the unknown response stage and correspond to the non-bidding state, bidding failure state, or bidding success state as the historical load training set. In some embodiments, the dataset is partitioned according to Table 1:

[0099] Table 1

[0100]

[0101] In some embodiments, step 220 involves calling the initial model to predict the historical load training set based on the candidate feature set to obtain the prediction error, including:

[0102] The candidate feature set is input into the initial model, triggering the forward propagation logic of the decision tree at the bottom layer of the initial model, and outputting the corresponding load prediction value. The specific calculation formula for the load prediction value output by the initial model is as follows:

[0103]

[0104] in, Let M be the predicted load value for the nth sample, M be the total number of decision trees, and L be the number of leaf nodes. Let B be the prediction weight of the l-th leaf node of the m-th decision tree, and let B be a Boolean function. For the input features of the nth sample, It is the feature subset corresponding to the l-th leaf node of the m-th decision tree.

[0105] During the prediction process, the initial model internally calculates the overall prediction loss function, the formula of which is:

[0106]

[0107] in, Here, N is the prediction loss function, and N is the total number of samples. Let be the actual load output of the nth sample, β be the regularization coefficient, and γ be the penalty coefficient for the leaf nodes. Compared with the traditional unregularized loss function, this formula introduces a penalty term for the tree structure complexity, effectively preventing model overfitting.

[0108] Furthermore, by comparing the predicted load values ​​with the actual load output in the historical load training set, the prediction error is calculated. The prediction error can be objectively quantified using either the relative root mean square error (RRMSE) or the mean absolute percentage error (MAPE). In some embodiments, the formulas for calculating RRMSE and MAPE are as follows:

[0109]

[0110]

[0111] In one embodiment, the initial model is an ensemble tree model. Step 220 extracts the gain and coverage metrics of each feature in the candidate feature set, including steps 222 to 226. Wherein:

[0112] Step 222: Extract the first and second derivative data of the prediction loss function during the prediction process.

[0113] The prediction loss function can be a mathematical function that measures the difference between the model output and the actual load data, such as the mean squared error function or the cross-entropy loss function. The first derivative data can be the gradient information of the loss function with respect to the model's predicted values. The second derivative data can be the Hessian matrix information of the loss function with respect to the model's predicted values.

[0114] For example, during the iterative training of the ensemble tree model, the prediction loss function of the current iteration step is expanded using a second-order Taylor expansion, and the first-order and second-order derivative terms in the Taylor polynomial are extracted.

[0115] In some embodiments, the second-order Taylor expansion approximation formula for the prediction loss function in the τth iteration is:

[0116]

[0117] In the formula, τ is the number of iterations, and N is the total number of samples. For first derivative data, For second derivative data, For decision tree functions, For the input features of the nth sample, Let be the feature subset corresponding to the l-th leaf node, β be the regularization coefficient, L be the number of leaf nodes, and γ be the penalty coefficient for the leaf nodes. This is used to predict weights. Compared to the conventional loss function calculation that relies solely on first-order gradient descent, this formula introduces second-order derivative data for Taylor expansion, preserving more information about the local loss surface and improving the accuracy of algorithm convergence and approximate calculation.

[0118] Step 224: Based on the second derivative data, calculate the influence of node samples when splitting features, as a coverage indicator.

[0119] The influence of a node sample can be the sum of the weights of the number of samples covered by the leaf nodes that split using the current feature, such as the sum of the weights of the left node of the feature split or the sum of the weights of the right node of the feature split.

[0120] For example, the second derivative data of all samples involved in the current feature splitting at the node are summarized, and the summed second derivative data are used to obtain the coverage index. In some embodiments, the formula for calculating the coverage index is:

[0121]

[0122] In the formula, Let H be the coverage index of the k-th feature, and H be the candidate feature set. These are samples belonging to the candidate feature set. The extracted second derivative data.

[0123] Step 226: Based on the first-order derivative data and the second-order derivative data, calculate the loss reduction caused by feature splitting as a gain index.

[0124] The loss reduction can be the amount by which the overall loss function decreases after the feature splits the current node into multiple child nodes, such as the difference between the loss before splitting and the loss of the left child node after splitting, or the difference between the loss before splitting and the loss of the right child node after splitting.

[0125] For example, a node is split into a left child node and a right child node according to its features. The ratio of the sum of squares of the first derivative data to the sum of the second derivative data in the left child node and the right child node is calculated respectively. The corresponding ratio of the parent node before splitting is subtracted to obtain the gain index.

[0126] In some embodiments, the formula for calculating the gain index is:

[0127]

[0128] In the formula, Let be the gain index for the k-th feature. and Let H be the feature subsets formed by the left and right leaf nodes after splitting the candidate feature set, respectively, and let H be the candidate feature set before splitting. For first derivative data, The data represents the second derivative, γ is the penalty coefficient for leaf nodes, and β is the regularization coefficient. Compared to conventional formulas that use principal component analysis linear combinations or a single Gini index to evaluate importance indicators, this calculation logic directly calls the derivative data at the bottom layer of the prediction model, directly binding the evaluation benefit of features to the decrease in the model's loss function, thus avoiding the loss of the physical meaning of features.

[0129] In this embodiment, by extracting the first and second derivative data of the prediction loss function during the prediction process, and calculating the node sample influence as a coverage index based on the second derivative data, and calculating the loss reduction as a gain index based on the first and second derivative data, gradient statistics based on the dimensionality reduction expression of complex model loss are realized. In addition, by introducing the second derivative data into the calculation formula of the feature split loss reduction, the local surface curvature information of the prediction loss function in the current feature space is better preserved, realizing direct data mapping between feature selection and prediction model optimization. This solves the technical problems of traditional single-filter evaluation algorithms being divorced from the actual operation mechanism of the model and relying solely on the first gradient leading to inaccurate feature evaluation, significantly reducing the computational overhead in the feature selection process and improving the robustness of the evaluation results.

[0130] In some embodiments, step 220 determines the importance of each feature based on gain and coverage metrics, including:

[0131] The acquired gain and coverage metrics are normalized to unify them to the same numerical magnitude, avoiding evaluation distortion caused by excessively large single absolute values. Subsequently, according to the pre-configured weights, a weighted average is calculated on the normalized gain and coverage metrics to obtain the final importance score for each candidate feature.

[0132] In some embodiments, the formula for calculating the importance index is:

[0133]

[0134] in, Let w be the importance score for the k-th candidate feature, where K is the total number of candidate features and w is the weight assigned to the indicator. Let be the gain index for the k-th candidate feature. Let be the coverage index for the k-th candidate feature. Compared with traditional methods that rely solely on gain or variance for evaluation, this method introduces index allocation weights and normalization calculation logic into the formula. By placing the gain index, which represents the local prediction contribution, and the coverage index, which represents the generalization ability of the global samples, on the same data scale for cross-weighting, a dynamic balance of multi-dimensional features is achieved. This solves the problem in traditional models where relying solely on information gain leads to feature selection easily getting trapped in local optima or overfitting specific samples, and significantly improves the stability of feature evaluation results.

[0135] In one embodiment, step 230 combines prediction error and importance index to iteratively eliminate candidate feature sets to obtain the optimal feature subset, including steps 232 to 236. Wherein:

[0136] Step 232: Sort the features in the candidate feature set in ascending order according to the importance index.

[0137] The importance metric can be a quantitative value that measures the contribution of each feature to the prediction model, such as a score calculated based on information gain or a score calculated based on coverage. The candidate feature set can be the collection of all features to be screened. Ascending sorting can be a data processing operation that arranges the data according to a rule of ascending numerical values.

[0138] For example, the importance score of each calculated feature is read, and a sorting algorithm is called to rearrange all candidate features in ascending order of importance score, so that the feature with the lowest importance score is at the end of the sequence and the feature with the highest importance score is at the beginning of the sequence.

[0139] Step 234: Remove the features at the end of the sorting in sequence, and obtain the prediction error re-output by the initial model after removing the features.

[0140] In this ranking, the feature at the bottom of the list can be a redundant feature with the lowest importance score and the smallest contribution to the model. Prediction error can be an objective quantitative value of the difference between the model's predicted value and the true value, such as root mean square error or mean absolute percentage error.

[0141] For example, one or more features at the end of the ascending sorted feature sequence are removed from the candidate feature set. Using the remaining feature set after feature removal, the initial model is re-triggered to perform a complete prediction operation on the historical load training set, and the newly generated prediction error value is extracted after the operation.

[0142] In some embodiments, during the t-th iteration of the removal operation, the feature with the lowest importance score is selected as the current removal target from the remaining feature set after ascending sorting. The logical expression for the feature removal action is:

[0143]

[0144]

[0145]

[0146] in, For features that are locked and removed in the t-th iteration, For importance indicators, This is the remaining feature subset before the t-th iteration. This is the updated subset of remaining features after the removal operation. This is the subset of features removed before the t-th iteration. This is the updated subset of features to be removed after the removal operation. Features with a maximum relative root mean square error or a minimum importance index are extracted as the removal targets for the current iteration, and the numerical change of the prediction error is forcibly monitored after each removal. After the set is updated, the updated remaining feature subset is used to re-trigger the initial model for a complete round of prediction calculation, and the newly generated prediction error value after the calculation is extracted for subsequent loop termination condition determination. By establishing a direct feedback loop between the step-by-step removal of features and the fluctuation of the overall model error, the problem of easily deleting core physical features during the dimensionality reduction of high-dimensional data is solved, ensuring that each dimensionality reduction action does not sacrifice the model's accuracy in predicting the true building load.

[0147] Step 236: If the prediction error is detected to no longer decrease or the number of remaining features is less than the preset constraint value, stop the elimination and determine the remaining features as the optimal feature subset.

[0148] The preset constraint value can be a pre-configured lower limit for the number of features allowed to be retained, such as 7 or 10 features. The optimal feature subset can be the combination of features retained after multiple rounds of screening that plays a decisive role in the baseline load forecast.

[0149] For example, after each feature removal and acquisition of a new prediction error, the new prediction error is numerically compared with the prediction error from the previous iteration. If, within a set number of consecutive iterations, the new prediction error is greater than or equal to the previous prediction error, or if the total number of features not yet removed is less than a preset constraint value, the iterative removal loop is interrupted and exited. All feature combinations remaining at the moment the loop stops are packaged and stored as the optimal feature subset. In some embodiments, if it is detected that the prediction error continues to decrease and the number of remaining features is greater than or equal to a preset constraint value, the feature removal and model re-prediction actions continue.

[0150] In this embodiment, by sorting features in ascending order based on importance indicators, the last features are removed sequentially, and the prediction error re-output by the model is obtained. The removal process stops when the prediction error no longer decreases or the number of remaining features reaches the lower limit. This achieves dynamic adaptive interaction between the feature selection process and the actual prediction accuracy of the model, thereby overcoming the shortcomings of traditional filtering feature selection that cannot take into account the underlying logic of the model. It ensures the accuracy of model prediction while removing redundant and noisy data, and effectively eliminates the logical loopholes of the algorithm falling into excessive removal or infinite looping.

[0151] In some embodiments, the results of the selection of air conditioning baseline load characteristics for government office buildings, commercial office buildings, shopping malls, and large supermarkets after the aforementioned feature selection are shown in Table 2:

[0152] Table 2

[0153]

[0154] According to records, the baseline load forecast accuracy of air conditioning for these four types of buildings in a certain location during a week in August 2023 met the requirements of the electricity market demand response, with RRMSE all less than 11.2% and MAPE all less than 10.8%. Compared with deep learning algorithms such as Long Short-Term Memory Networks and Convolutional Networks, the accuracy was improved by more than 5%; the model training time was 32 seconds, which was reduced by more than 32% compared with deep learning algorithms.

[0155] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0156] Based on the same inventive concept, this application also provides a building baseline load forecasting model training system 400 for implementing the building baseline load forecasting model training method described above. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the building baseline load forecasting model training system 400 provided below can be found in the limitations of the building baseline load forecasting model training method described above, and will not be repeated here.

[0157] In one exemplary embodiment, such as Figure 4 As shown, a building baseline load prediction model training system 400 is provided, including an acquisition module 401, a processing module 403, a rejection module 405, and a training module 407, wherein:

[0158] The acquisition module 401 is used to acquire the candidate feature set and historical load training set of the target building.

[0159] The processing module 403 is used to call the initial model to predict the historical load training set based on the candidate feature set, obtain the prediction error, extract the gain index and coverage index of each feature in the candidate feature set, and determine the importance index of each feature based on the gain index and coverage index.

[0160] The elimination module 405 is used to iteratively eliminate candidate feature sets by combining prediction error and importance index to obtain the optimal feature subset.

[0161] Training module 407 is used to input the optimal feature subset and historical load training set into the initial model for training to obtain the target model. The target model is used to receive online feature data of the target building, extract the target data corresponding to the optimal feature subset from the online feature data for prediction, and obtain the baseline load prediction result.

[0162] In one embodiment, the acquisition module 401 acquires a candidate feature set and a historical load training set for the target building, including:

[0163] Obtain the initial load data of the target building during the historical demand response process, as well as the time status and user participation status corresponding to the initial load data;

[0164] The initial load data is classified according to the time status and user participation status, resulting in multiple load data subsets;

[0165] Extract the load data subsets that are in the target state from multiple load data subsets and use them as the historical load training set.

[0166] In one embodiment, the time state includes an unknown response phase, a known response phase, and an execution response phase;

[0167] User participation status includes no bidding, failed bidding, and successful bidding.

[0168] Extract the load data subsets that are in the target state from multiple load data subsets to serve as the historical load training set, including:

[0169] A subset of load data that is in the unknown response phase and corresponds to the non-bid, bid failure, or bid success states will be used as the historical load training set.

[0170] In one embodiment, the initial model is an ensemble tree model;

[0171] Processing module 403 extracts the gain and coverage metrics of each feature in the candidate feature set, including:

[0172] During the prediction process, the first and second derivative data of the prediction loss function are extracted;

[0173] Based on second derivative data, the influence of node samples when features are split is calculated as a coverage indicator.

[0174] Based on first-order and second-order derivative data, the loss reduction caused by feature splitting is calculated and used as a gain index.

[0175] In one embodiment, the elimination module 405 combines prediction error and importance index to iteratively eliminate candidate feature sets to obtain an optimal feature subset, including:

[0176] Sort the features in the candidate feature set in ascending order based on their importance index;

[0177] Remove the features at the end of the sorting list one by one, and obtain the prediction error of the initial model after removing the features;

[0178] If the prediction error no longer decreases or the number of remaining features is less than the preset constraint value, the elimination process stops, and the remaining features are determined as the optimal feature subset.

[0179] Each module in the aforementioned building baseline load prediction model training system 400 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0180] In one exemplary embodiment, such as Figure 5 As shown, a load forecasting method is provided, which can be applied to... Figure 1 The following steps are used as an example to illustrate the process of data processing device 120, including steps 510 to 550. Wherein:

[0181] Step 510: Obtain online characteristic data of the target building during the demand response period.

[0182] The online characteristic data can be real-time environmental parameters or real-time operating status parameters that affect building electricity consumption, such as real-time outdoor temperature or real-time lighting power. The demand response period can be a specific time period during which the power grid dispatch center issues load adjustment instructions, such as peak shaving response period or valley filling response period.

[0183] For example, real-time data packets continuously uploaded by data acquisition devices 110 deployed inside and outside the target building are received in real time through a communication network. The real-time data packets are parsed and the corresponding values ​​are extracted and temporarily loaded into memory as online feature data.

[0184] Step 520: Extract the target data corresponding to the optimal feature from the online feature data.

[0185] The optimal feature can be the core feature retained after importance ranking and iterative elimination during the offline training phase, such as retained temperature features or historical load features. The target data can be a subset of online data that strictly corresponds to the data dimension of the optimal feature.

[0186] For example, the dimension labels of the optimal features pre-stored locally are read, the obtained online feature data is traversed, and the data is matched and filtered according to the dimension labels. The successfully matched data fields are copied separately and concatenated into a data vector in a standard format as the target data.

[0187] Step 530: Input the target data into the pre-trained target model for prediction to obtain the baseline load prediction result of the target building.

[0188] The target model can be an ensemble tree model that has completed parameter iterative updates and has high-precision prediction capabilities, such as the LightGBM model trained offline. The baseline load forecast result can be the expected power consumption value under the assumption that the target building does not participate in load regulation.

[0189] For example, the target model pre-loaded into memory is invoked, and the concatenated target data is passed as input parameters to the input interface of the target model. This triggers the forward propagation of the decision tree and the calculation of leaf node weights at the bottom layer of the target model, outputting the corresponding expected power consumption value, which is then determined as the baseline load prediction result.

[0190] Step 540: Obtain the actual operating load of the target building during the demand response period.

[0191] The actual operating load can be the actual electrical power consumed by the target building during the demand response period, such as the real-time active power of the main line or the real-time power consumption of the central air conditioning system.

[0192] For example, the measurement readings fed back by the smart meter or energy consumption control equipment 130 arranged in the power distribution room of the target building are read in real time through the metering interface. After unit conversion and outlier filtering of the measurement readings, the objective actual operating load is obtained.

[0193] Step 550: Based on the baseline load forecast results and the actual operating load, determine the load reduction amount for the target building, and generate the corresponding demand response scheduling instruction according to the load reduction amount.

[0194] The optimal features and target model are obtained through the training method of the baseline load forecasting model as described above. The load reduction amount can be a quantified value of the difference between the baseline load forecast result and the actual power consumption. The demand response scheduling instruction can be a machine control message used to directly control the operating logic of the underlying physical equipment, such as a temperature increase instruction or a host power limit instruction.

[0195] For example, the difference between the baseline load forecast and the actual operating load is calculated, and this difference is determined as the load reduction amount. Then, a pre-configured target reduction threshold is retrieved, and the load reduction amount is compared with the target reduction threshold. If the load reduction amount is determined to be less than the target reduction threshold, the required voltage reduction of the equipment operating power is calculated based on the difference between the target reduction threshold and the load reduction amount. A demand response scheduling instruction for reducing the operating parameters of the target energy-consuming equipment is generated and sent to the underlying energy consumption control device 130 via the communication network. In some embodiments, if the load reduction amount is determined to be greater than or equal to the target reduction threshold, a scheduling instruction to maintain the current operating state is generated.

[0196] In this embodiment, by acquiring online feature data and extracting target data with corresponding optimal features, the baseline load forecast result is obtained using the target model. Based on the baseline load forecast result and the actual operating load, the load reduction amount is calculated, and then the corresponding demand response scheduling instruction is generated. This realizes the precise control of physical energy-consuming equipment from accurate online data forecasting. At the engineering level, it provides an action basis for power grid dispatching, effectively avoids equipment downtime caused by excessive load reduction, and greatly improves the stability and reliability of demand response regulation of the new power system.

[0197] In one embodiment, step 550 generates a corresponding demand response scheduling instruction based on the load reduction amount, including steps 552 to 556. Wherein:

[0198] Step 552: Obtain the target reduction threshold for the target building during the demand response period.

[0199] The target reduction threshold can be a numerical value of the expected reduction in electricity consumption of a building issued by the power grid dispatch center, such as a fixed power difference or a percentage power ratio.

[0200] For example, demand response event messages issued by the power grid dispatch center are received in real time through the communication network, the message content is parsed, the specific power reduction value allocated to the target building is extracted, and it is temporarily loaded into memory as the target reduction threshold.

[0201] Step 554: If the load reduction amount is determined to be less than the target reduction threshold, a demand response scheduling instruction is generated based on the difference between the target reduction threshold and the load reduction amount to reduce the operating power of the target energy-consuming equipment.

[0202] The difference can be the power shortfall between the target reduction threshold and the currently calculated load reduction amount. The target energy-consuming equipment can be physical electrical facilities within the building whose power consumption can be dynamically adjusted, such as central air conditioning units or building lighting networks.

[0203] For example, an objective numerical comparison is made between the load reduction amount and the target reduction threshold. If the load reduction amount is found to be less than the target reduction threshold, the target reduction threshold is subtracted from the load reduction amount to calculate the power difference that needs to be further reduced. Based on the power difference and in conjunction with the pre-configured equipment energy efficiency mapping relationship, the operating parameters that the target energy-consuming equipment needs to be reduced are calculated. The operating parameters are then encapsulated into machine-readable control messages and sent out as demand response scheduling instructions.

[0204] Step 556: If the load reduction amount is determined to be greater than or equal to the target reduction threshold, a scheduling instruction is generated to maintain the current operating status of the target energy-consuming equipment.

[0205] The current operating status can be the real-time power level or hardware setting parameters of the target energy-consuming equipment at the time of load determination, such as the current set temperature or the current compressor operating frequency.

[0206] For example, when the load reduction amount is found to have reached or exceeded the target reduction threshold after comparing the numerical values, a machine control message containing an unchanged identifier is directly generated. This message is then sent as a scheduling instruction to the underlying energy consumption control device 130, so that the target energy consumption device continues to operate according to the original operating parameters and no longer triggers additional power reduction actions.

[0207] In this embodiment, by obtaining the target load reduction threshold and generating an instruction to reduce the operating power based on the power difference when the load reduction amount does not meet the target, and generating an instruction to maintain the current operating state when the load reduction amount has met the target, precise two-way closed-loop control of the operating state of the target building's physical energy-consuming equipment is achieved. This avoids response default caused by insufficient reduction or abnormal shutdown of physical equipment caused by excessive reduction, and further strengthens the safety and reliability of the entire demand response process of the new power system.

[0208] Figure 6 This is a schematic diagram illustrating the interaction timing between building baseline load prediction training and predictive control, provided as an embodiment. Figure 6 As shown, the collaborative interaction process of multiple devices may include:

[0209] During the offline feature selection and model training phase, candidate feature sets and historical load training sets are uploaded to data processing device 120 via data acquisition device 110. Inside data processing device 120, the initial model is invoked for prediction and gain and coverage metrics are extracted. Iterative elimination is performed by combining prediction error and importance metrics to obtain the optimal features. The target model is then trained using the optimal features and the historical load training set.

[0210] During the online prediction and control phase, the data acquisition device 110 sends online characteristic data and actual operating load to the data processing device 120 in real time. Within the data processing device 120, target data corresponding to the optimal features is extracted and input into the target model to predict the baseline load forecast. Further, the data processing device 120 calculates the difference between the baseline load forecast and the actual operating load to determine the load reduction amount, and generates corresponding demand response scheduling instructions, such as power reduction instructions or status quo maintenance instructions, based on the comparison between the load reduction amount and the target reduction threshold. Subsequently, the data processing device 120 issues demand response scheduling instructions to the energy consumption control device 130. The energy consumption control device 130 receives the demand response scheduling instructions and adjusts the operating status of target energy-consuming equipment, such as central air conditioning compressors or building lighting circuits, according to the instructions.

[0211] In this embodiment, through multi-node time-series interaction between data acquisition device 110, data processing device 120, and energy consumption control device 130, feature screening and model training are completed in the offline stage, and real-time data prediction and instruction issuance are completed in the online stage. By collecting underlying data, performing algorithm prediction on the edge side, and controlling terminal physical devices, the timeliness of data generated by demand response scheduling instructions and the accuracy of physical control are ensured, thereby improving the overall execution efficiency and reliability of building nodes participating in the demand response of the new power system.

[0212] Based on the same inventive concept, this application also provides a load forecasting system 700 for implementing the load forecasting method described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more embodiments of the load forecasting system 700 provided below can be found in the limitations of the load forecasting method described above, and will not be repeated here.

[0213] In one exemplary embodiment, such as Figure 7 As shown, a load forecasting system 700 is provided, comprising a feature acquisition module 701, a feature extraction module 703, a load forecasting module 705, a load acquisition module 707, and a schedule generation module 709, wherein:

[0214] The feature acquisition module 701 is used to acquire online feature data of the target building during the demand response period.

[0215] The feature extraction module 703 is used to extract the target data corresponding to the optimal feature from the online feature data.

[0216] The load forecasting module 705 is used to input target data into a pre-trained target model for forecasting, and obtain the baseline load forecasting results of the target building.

[0217] The load acquisition module 707 is used to acquire the actual operating load of the target building during demand response.

[0218] The scheduling generation module 709 is used to determine the load reduction amount of the target building based on the baseline load forecast results and the actual operating load, and generate corresponding demand response scheduling instructions according to the load reduction amount.

[0219] The optimal features and target model are obtained through the training method of the building baseline load prediction model as described in any of the preceding items.

[0220] In one embodiment, the scheduling generation module 709 generates a corresponding demand response scheduling instruction based on the load reduction amount, including:

[0221] Obtain the target reduction threshold for the target building during the demand response period;

[0222] If the load reduction amount is determined to be less than the target reduction threshold, a demand response scheduling instruction is generated based on the difference between the target reduction threshold and the load reduction amount to reduce the operating power of the target energy-consuming equipment.

[0223] If the load reduction amount is determined to be greater than or equal to the target reduction threshold, a scheduling instruction is generated to maintain the current operating status of the target energy-consuming equipment.

[0224] Each module in the aforementioned load forecasting system 700 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0225] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 8 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores training data for the building baseline load forecasting model and load forecasting data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When executed by the processor, the computer program can implement a building baseline load forecasting model training method or a load forecasting method.

[0226] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0227] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0228] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0229] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0230] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0231] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0232] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0233] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for training a building baseline load prediction model, characterized in that, The method includes: Obtain the candidate feature set and historical load training set of the target building; The initial model is invoked to predict the historical load training set based on the candidate feature set to obtain the prediction error. The gain index and coverage index of each feature in the candidate feature set are extracted. Based on the gain index and the coverage index, the importance index of each feature is determined. By combining the prediction error and the importance index, the candidate feature set is iteratively eliminated to obtain the optimal feature subset; The optimal feature subset and the historical load training set are input into the initial model for training to obtain the target model; the target model is used to receive online feature data of the target building, extract the target data corresponding to the optimal feature subset from the online feature data for prediction, and obtain the baseline load prediction result.

2. The method according to claim 1, characterized in that, The acquisition of the candidate feature set and historical load training set of the target building includes: Obtain the initial load data of the target building during the historical demand response process, as well as the time status and user participation status corresponding to the initial load data; Based on the time status and the user participation status, the initial load data is classified to obtain multiple load data subsets; Extract the load data subset that is in the target state from the multiple load data subsets, and use it as the historical load training set.

3. The method according to claim 2, characterized in that, The time state includes an unknown response phase, a known response phase, and an execution response phase; The user participation status includes no bidding status, failed bidding status, and successful bidding status; The step of extracting the load data subset in the target state from the multiple load data subsets as the historical load training set includes: The subset of load data that is in the unknown response stage and corresponds to the non-bid, bid failure, or bid success state will be used as the historical load training set.

4. The method according to claim 1, characterized in that, The initial model is an ensemble tree model; The extraction of gain and coverage metrics for each feature in the candidate feature set includes: During the prediction process, the first and second derivative data of the prediction loss function are extracted; Based on the second derivative data, the influence of node samples when the feature is split is calculated and used as the coverage index. Based on the first-order derivative data and the second-order derivative data, the loss reduction caused by the feature split is calculated and used as the gain index.

5. The method according to claim 1, characterized in that, The process of combining the prediction error and the importance index to iteratively eliminate candidate feature sets to obtain an optimal feature subset includes: Sort the features in the candidate feature set in ascending order according to the importance index; The features at the end of the sort are removed sequentially, and the prediction error re-output by the initial model after removing the features is obtained. If the prediction error is detected to no longer decrease or the number of remaining features is less than a preset constraint value, the elimination process stops, and the remaining features are determined as the optimal feature subset.

6. A training system for a building baseline load prediction model, characterized in that, include: The acquisition module is used to acquire the candidate feature set and historical load training set of the target building; The processing module is used to call the initial model to predict the historical load training set based on the candidate feature set, obtain the prediction error, extract the gain index and coverage index of each feature in the candidate feature set, and determine the importance index of each feature based on the gain index and the coverage index. The elimination module is used to combine the prediction error and the importance index to iteratively eliminate the candidate feature set to obtain the optimal feature subset; The training module is used to input the optimal feature subset and the historical load training set into the initial model for training to obtain the target model; the target model is used to receive the online feature data of the target building, extract the target data corresponding to the optimal feature subset from the online feature data for prediction, and obtain the baseline load prediction result.

7. A load forecasting method, characterized in that, include: Acquire online characteristic data of the target building during demand response; Extract the target data corresponding to the optimal feature from the online feature data; The target data is input into a pre-trained target model for prediction to obtain the baseline load prediction result of the target building; Obtain the actual operating load of the target building during the demand response period; Based on the baseline load forecast results and the actual operating load, the load reduction amount for the target building is determined, and a corresponding demand response scheduling instruction is generated according to the load reduction amount. The optimal features and the target model are obtained by the training method of the baseline load prediction model as described in any one of claims 1 to 5.

8. The method according to claim 7, characterized in that, Generate corresponding demand response scheduling instructions based on the load reduction amount, including: Obtain the target reduction threshold for the target building during the demand response period; If it is determined that the load reduction amount is less than the target reduction threshold, a demand response scheduling instruction is generated based on the difference between the target reduction threshold and the load reduction amount to reduce the operating power of the target energy-consuming equipment. If the load reduction amount is determined to be greater than or equal to the target reduction threshold, a scheduling instruction is generated to maintain the current operating state of the target energy-consuming equipment.

9. A load forecasting system, characterized in that, The system includes: The feature acquisition module is used to acquire online feature data of the target building during the demand response period; The feature extraction module is used to extract the target data corresponding to the optimal feature from the online feature data; The load forecasting module is used to input the target data into a pre-trained target model for forecasting, and obtain the baseline load forecasting result of the target building; The load acquisition module is used to acquire the actual operating load of the target building during the demand response period; The scheduling generation module is used to determine the load reduction amount of the target building based on the baseline load forecast results and the actual operating load, and generate corresponding demand response scheduling instructions according to the load reduction amount. The optimal features and the target model are obtained by the training method of the building baseline load prediction model as described in any one of claims 1 to 5.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the building baseline load forecasting model training method according to any one of claims 1 to 5, or the steps of the load forecasting method according to any one of claims 7 to 8.

Images