A building load prediction method, device, equipment and medium

Abstract

The application discloses a building load prediction method, device, equipment and medium, and belongs to the electric energy prediction field.The method is as follows: first load data of a target building and second load data of a plurality of same-type adjacent buildings corresponding to the target building are collected; the first load data and each second load data are input into a black box adaptive network which has been pre-trained, so as to respectively obtain an input embedding vector, a space embedding vector and a feature embedding vector; the input embedding vector, the space embedding vector and the feature embedding vector are input into a preset large language model, so as to obtain a load prediction embedding vector representing a load prediction result; wherein the large language model is iteratively trained on the black box adaptive network, so as to avoid accessing internal parameters of the large language model; and the load prediction result of the target building is determined based on the load prediction embedding vector.The application can improve the building load prediction accuracy in a few-sample or zero-sample scenario.

Description

Technical Field

[0001] This application relates to the field of power forecasting, and more particularly to a method, apparatus, equipment and medium for predicting building loads. Background Technology

[0002] With the widespread adoption of advanced metering infrastructure, short-term building-level load forecasting has become a core component of modern power system operation and management. Accurate load forecasting not only helps power operators optimize energy dispatch and ensure stable grid operation, but also assists end-users in reducing electricity costs and provides precise data support for various decisions in the electricity market. Therefore, how to utilize load data to build high-performance forecasting models has become a research hotspot in the field of smart grids.

[0003] Existing technologies face the following bottlenecks: While machine learning methods are better at extracting complex spatiotemporal features than traditional statistical methods, they rely on a large amount of load data. Data gaps in newly established communities and equipment failures often lead to few-sample / zero-sample scenarios, limiting the effectiveness. Few-sample transfer learning requires similar source domain data; otherwise, it is prone to "negative transfer." Moreover, source data is difficult to obtain due to privacy protection and cannot cover zero samples, resulting in poor generalization ability and inability to accurately predict loads in zero-sample scenarios. Although large language models have strong generalization and few-sample / zero-sample advantages, the high memory overhead during training, the lack of public model parameters, and the inability of existing large language models to consider load spatial correlations on their own result in low accuracy of the final load prediction results. Summary of the Invention

[0004] This application provides a building load forecasting method, apparatus, equipment, and medium that can solve the problem of low accuracy of load forecasting results in scenarios with few or no samples in the prior art.

[0005] Some embodiments of this application provide a building load forecasting method, including: Collect the first load data of the target building and the second load data of several adjacent buildings of the same type corresponding to the target building; The first load data and each of the second load data are input into the pre-trained black-box adaptation network to obtain the input embedding vector characterizing the electricity consumption time series pattern of the target building, the spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and the feature embedding vector characterizing the load dynamic mode of the target building. The input embedding vector, the spatial embedding vector, and the feature embedding vector are input into a preset large language model to obtain a load prediction embedding vector representing the load prediction result; wherein, the large language model is trained iteratively on the black-box adaptation network to avoid access to the internal parameters of the large language model; The load prediction result for the target building is determined based on the load prediction embedding vector.

[0006] Compared to existing technologies, the above embodiments have the following advantages: By simultaneously introducing the load data of the target building itself and the load data of adjacent buildings of the same type, the effective information sources for the target building under conditions of insufficient samples are explicitly supplemented at the data level; and a black-box adaptation network is used to uniformly model the above multi-source load data, extracting multi-dimensional spatiotemporal features that take into account temporal evolution patterns, spatial synergistic effects, and load dynamic modes, thereby enhancing the ability to characterize the intrinsic mechanism of building load at the feature level. On this basis, the multi-dimensional spatiotemporal features are mapped to the input space of a large language model, enabling the large language model to leverage its generalization advantages in few-sample and zero-sample inference without directly perceiving the original load data and spatial structure. At the same time, iterative training through a black-box adaptation network avoids access restrictions on the internal parameters of the large language model, significantly reducing the difficulty of model adaptation and computational overhead. Finally, the prediction result is mapped through an independently trained output layer, enabling the overall method to effectively improve the accuracy of building load prediction in few-sample or zero-sample scenarios while ensuring feasibility.

[0007] Further, the black-box adaptation network includes: an input embedding layer, a spatial embedding layer, and a feature extraction layer; the input of the first load data and each of the second load data into the pre-trained black-box adaptation network yields an input embedding vector characterizing the electricity consumption time-series pattern of the target building, a spatial embedding vector characterizing the synergistic effect between the target building and its adjacent buildings of the same type, and a feature embedding vector characterizing the load dynamic mode of the target building, including: The first load data and each of the second load data are input into the pre-trained black-box adaptation network, so that the first load data is normalized and linearly transformed through the input embedding layer to generate the input embedding vector. The first load data and each of the second load data are input into the pre-trained black-box adaptation network. The first load data and the dynamically sampled second load data are normalized and aggregated through the spatial embedding layer to obtain aggregated feature data. The aggregated feature data is then linearly transformed to generate the spatial embedding vector. The first load data and each of the second load data are input into the pre-trained black-box adaptation network. The first load data is constructed into a multidimensional augmented data matrix by using the Hankel operator through the feature extraction layer. The dominant feature value is extracted by dynamic pattern decomposition of the multidimensional augmented data matrix, and the feature embedding vector is generated by linear transformation based on the dominant feature value.

[0008] Compared to existing technologies, the above embodiments have the following beneficial effects: By dividing the black-box adaptation network into an input embedding layer, a spatial embedding layer, and a feature extraction layer, targeted modeling is performed on the time-series characteristics of load data, the spatial coordination relationship between buildings, and the load dynamics mode, respectively, so that load information from different sources and with different physical meanings is structurally decoupled and uniformly mapped. On the one hand, it avoids the semantic mismatch problem caused by directly inputting the original multi-source load data into the large language model; on the other hand, by explicitly constraining the expression boundaries of various features in the form of embedding vectors, the large language model can still make full use of its reasoning and generalization capabilities without being aware of the underlying load mechanism, thereby improving the overall stability and accuracy of load prediction in scenarios with few or no samples.

[0009] Further, the step of constructing a multidimensional augmented data matrix from the first load data using the Hankel operator, and then performing dynamic pattern decomposition on the multidimensional augmented data matrix to extract dominant feature values, includes: The first load data is sampled using a sliding window of a set length to construct the multidimensional augmented data matrix that captures local dynamic dependencies; Linear operators are constructed based on the mapping relationship of continuous time steps in the multidimensional augmented data matrix to describe the dynamic evolution of the load. Singular value decomposition and eigenvalue decomposition are performed on the linear operator to obtain the dominant eigenvalues ​​that reflect the load change trend and oscillation characteristics.

[0010] Compared to existing technologies, the above embodiments have the following advantages: By introducing the Hankel operator and dynamic mode decomposition, the load sequence is augmented in multiple dimensions and the dominant feature values ​​are extracted, transforming the originally highly non-stationary and noisy time series into a dynamic feature space that can be characterized by a small number of dominant modes. This processing method can effectively separate the long-term trend components and periodic oscillation components in load changes, reduce the interference of short-term random fluctuations on feature expression, and make the extracted feature embedding vectors closer to the intrinsic physical mechanism of load evolution, thus maintaining strong interpretability and predictive robustness even under conditions of insufficient samples.

[0011] Further, the step of normalizing and aggregating the first load data and the dynamically sampled second load data to obtain aggregated feature data includes: Based on a preset similarity measurement criterion, a set number of first buildings are randomly selected from the adjacent buildings of the same type. The first load data and the second load data corresponding to each of the first buildings are normalized to obtain a standard load sequence. The aggregated characteristic data, which characterizes the shared electricity consumption of the community, is obtained by averaging all the standard load sequences.

[0012] Compared to existing technologies, the above embodiments have the following advantages: By dynamically sampling from adjacent buildings of the same type based on similarity metrics, and performing dimensional normalization and mean aggregation on the load sequences of multiple buildings, aggregated feature data characterizing the common electricity consumption behavior of the community is constructed, effectively mitigating the uncertainty caused by the scarcity or abnormal fluctuations of load data for a single target building. This approach avoids the noise accumulation problem caused by fixing adjacency relationships or fully introducing external building data, enabling the spatial embedding vector to stably reflect the relative position and collaborative characteristics of the target building within the same type of building group, thereby improving the reliability of the model in cross-building generalization prediction.

[0013] Furthermore, the iterative training of the black-box adaptation network using the zero-order gradient through a stochastic approximation algorithm includes: Obtain the first parameter of the black-box adaptation network in the current iteration, generate a random perturbation vector using a preset Bernoulli distribution, and apply the random perturbation vector to the first parameter to obtain a pair of positive perturbation parameters and negative perturbation parameters; The positive perturbation parameters and the negative perturbation parameters are respectively input into the black box adaptation network, and the corresponding positive perturbation loss value and negative perturbation loss value are calculated by combining the inference results of the large language model. Based on the difference between the positive perturbation loss value and the negative perturbation loss value, and in conjunction with the random perturbation vector, the zeroth-order gradient characterizing the evolution direction of the first parameter is calculated. The black-box adaptation network is updated using the zero-order gradient obtained in each iteration, so that the black-box adaptation network tends to converge while avoiding access to the internal parameters of the large language model, thus obtaining the trained black-box adaptation network.

[0014] Compared to existing technologies, the above embodiments have the following advantages: By employing a zero-order gradient approximation algorithm based on random perturbation to train the black-box adaptation network, effective optimization of overall prediction performance is achieved without accessing or backpropagating the internal parameters of the large language model. On the one hand, this training method avoids the engineering limitations of unavailable parameters and high computational resource overhead of large language models, significantly reducing the difficulty of system deployment; on the other hand, by using the difference between positive and negative perturbation loss to estimate the parameter evolution direction, the black-box adaptation network can gradually learn embedding mapping relationships that better match the inference preferences of the large language model, thereby achieving efficient collaboration with the large language model.

[0015] Further, determining the load forecast result of the target building based on the load forecast embedding vector includes: The load prediction embedding vector is input into the pre-trained input layer, and the load prediction embedding vector is linearly transformed by the input layer to obtain the load prediction result; wherein, the input layer completes iterative training based on the first-order gradient of backpropagation during the pre-training stage.

[0016] Compared to existing technologies, the above embodiments have the following advantages: By setting up an independently pre-trained output mapping layer, the load prediction embedding vector is linearly mapped to the final load prediction result, effectively decoupling the semantic reasoning space and the numerical prediction space. This structure avoids the instability caused by directly relying on large language models to generate numerical results, and allows the numerical accuracy of the prediction results to be controlled and optimized through traditional supervised learning methods. Thus, while maintaining the generalization advantages of large language models, it ensures that the load prediction results meet the engineering requirements of power industry for accuracy and controllability.

[0017] Furthermore, the output layer completes iterative training based on the first-order gradient of backpropagation during the pre-training phase, including: The first load prediction result output by the output layer in the current iteration step is compared with the corresponding first real load data, and the prediction loss representing the prediction deviation is calculated. Based on the chain rule, the predicted loss is backpropagated along the output layer to calculate the first gradient at the current iteration step. The output layer is updated based on the first-order gradient obtained in each iteration by using a preset adaptive moment estimation optimizer to minimize the prediction loss.

[0018] Compared to existing technologies, the above embodiments have the following advantages: By pre-training the output layer based on first-order gradient backpropagation and an adaptive moment estimation optimizer, the prediction error can be minimized efficiently and stably. On the one hand, the adaptive optimization strategy can automatically adjust the learning rate according to different time scales and load amplitude changes, improving the training convergence speed; on the other hand, the independent optimization of the output layer allows the system to adapt to new scenarios by only fine-tuning the mapping layer when facing different buildings and different load levels, thereby further enhancing the practicality and scalability of the overall method in applications with few samples and across scenarios.

[0019] Another embodiment of this application provides a building load prediction device, including: a data acquisition module, a feature extraction module, a model calling module, and a mapping module; The acquisition module is used to acquire the first load data of the target building and the second load data of several adjacent buildings of the same type corresponding to the target building. The feature extraction module is used to input the first load data and each of the second load data into the pre-trained black-box adaptation network to obtain an input embedding vector characterizing the electricity consumption time sequence pattern of the target building, a spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and a feature embedding vector characterizing the load dynamic mode of the target building. The model invocation module is used to input the input embedding vector, the spatial embedding vector, and the feature embedding vector into a preset large language model to obtain a load prediction embedding vector representing the load prediction result; wherein, the large language model avoids access to the internal parameters of the large language model by iteratively training the black-box adaptation network. The mapping module is used to determine the load prediction result of the target building based on the load prediction embedding vector.

[0020] Another embodiment of this application also provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the steps of the building load prediction method of this application.

[0021] Another embodiment of this application also provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform the steps of the building load forecasting method of this application. Attached Figure Description

[0022] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a building load prediction method provided in some embodiments of this application; Figure 2 This is a schematic diagram of a load forecasting framework structure provided in some embodiments of this application; Figure 3 This is a structural schematic diagram of a building load prediction device provided in some embodiments of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0026] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0029] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0030] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0031] With the widespread adoption of advanced metering infrastructure, short-term building-level load forecasting has become a core component of modern power system operation and management. Accurate load forecasting not only helps power operators optimize energy dispatch and ensure stable grid operation, but also assists end-users in reducing electricity costs and provides precise data support for various decisions in the electricity market. Therefore, how to utilize load data to build high-performance forecasting models has become a research hotspot in the field of smart grids.

[0032] Existing technologies face the following bottlenecks: While machine learning methods are better at extracting complex spatiotemporal features than traditional statistical methods, they rely on a large amount of load data. Data gaps in newly established communities and equipment failures often lead to few-sample / zero-sample scenarios, limiting the effectiveness. Few-sample transfer learning requires similar source domain data; otherwise, it is prone to "negative transfer." Moreover, source data is difficult to obtain due to privacy protection and cannot cover zero samples, resulting in poor generalization ability and inability to accurately predict loads in zero-sample scenarios. Although large language models have strong generalization and few-sample / zero-sample advantages, the high memory overhead during training, the lack of public model parameters, and the inability of existing large language models to consider load spatial correlations on their own result in low accuracy of the final load prediction results.

[0033] Please refer to Figure 1 To address the issue of low accuracy in load forecasting under limited or zero-sample scenarios in existing technologies, this application provides a building load forecasting method, comprising the following steps S101 to S104: S101: Collect the first load data of the target building and the second load data of several adjacent buildings of the same type corresponding to the target building.

[0034] Since the building-level load forecasting problem aims to learn a model composed of trainable weights... Parameterized prediction function This function will pass the past The load data of the step is mapped to the next step. The load data is collected step by step. Therefore, when collecting the first load data and the second load data, it is for real-time measurement. Step load data to predict the future The load data is from a specific step. Furthermore, when the model is applied to load data from other scenarios, the number of buildings may differ. Therefore, It needs to be inductive so that it can predict load data for any number of buildings during real-time operation.

[0035] set up This represents the set of target buildings during the real-time operation phase, where the total number of target buildings is... ,Right now .set up Indicates the target building From time Time The first load data observed. Let... Indicates the target building A collection of adjacent buildings of the same type, and Indicates from time Time of The second load data for adjacent buildings of the same type. Then, in the future... The load data in the step can be predicted using the learned model as follows: ; in, For the target building From time arrive Forecasted load data; That is, the prediction model obtained through training. These are the optimal parameters for the prediction model. It is understandable that, referring to... Figure 2 The prediction model mentioned in the embodiments of this application includes: a black-box adaptation network (BlackInter, i.e., a black-box intermediate interface or black-box interaction layer), a pre-trained large language model (LLM), and an output layer (i.e., an output projection layer). The parameters that need to be trained include the trainable parameters in the black-box adaptation network and the output layer.

[0036] Preferably, in some embodiments of this application, all load data are a sequence of load values ​​for the corresponding building over several consecutive time steps.

[0037] S102: Input the first load data and each of the second load data into the pre-trained black-box adaptation network to obtain the input embedding vector characterizing the electricity consumption time sequence pattern of the target building, the spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and the feature embedding vector characterizing the load dynamic mode of the target building.

[0038] Furthermore, in some embodiments of this application, the black-box adaptation network includes: an input embedding layer, a spatial embedding layer, and a feature extraction layer.

[0039] Further, in some embodiments of this application, the step of inputting the first load data and each of the second load data into a pre-trained black-box adaptation network to obtain an input embedding vector characterizing the electricity consumption time-series pattern of the target building, a spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and a feature embedding vector characterizing the load dynamic mode of the target building, includes: The first load data and each of the second load data are input into the pre-trained black-box adaptation network, so that the first load data is normalized and linearly transformed through the input embedding layer to generate the input embedding vector. The first load data and each of the second load data are input into the pre-trained black-box adaptation network. The first load data and the dynamically sampled second load data are normalized and aggregated through the spatial embedding layer to obtain aggregated feature data. The aggregated feature data is then linearly transformed to generate the spatial embedding vector. The first load data and each of the second load data are input into the pre-trained black-box adaptation network. The first load data is constructed into a multidimensional augmented data matrix by using the Hankel operator through the feature extraction layer. The dominant feature value is extracted by dynamic pattern decomposition of the multidimensional augmented data matrix, and the feature embedding vector is generated by linear transformation based on the dominant feature value.

[0040] Preferably, in some embodiments of this application, the step of obtaining the input embedding vector includes: for the input first payload data First, the first load data is normalized using the mean and variance. Then, the normalized first load data is processed using a block-based method. Divide the data into several blocks. The advantage of data partitioning is that it preserves local semantic information by grouping local time series load data into a single block. Let... For block length, If the step size is the non-overlapping area between two consecutive blocks, then the number of blocks... It can be calculated in the following ways: ; in, This represents the floor function.

[0041] set up This represents the first load data after partitioning, see reference. Figure 2 Samples after normalization and block division It is fed into a linear layer, which can be represented as ; in, and These are the learnable parameters input into the embedding layer. It is the generated input embedding vector, and It is the column size of the input to the pre-trained language model (LLM).

[0042] This application divides the black-box adaptation network into an input embedding layer, a spatial embedding layer, and a feature extraction layer, respectively, to model the temporal series characteristics of load data, the spatial coordination relationships between buildings, and the load dynamics modes. This allows load information from different sources and with different physical meanings to be structurally decoupled and uniformly mapped. On the one hand, it avoids the semantic mismatch problem caused by directly inputting raw multi-source load data into a large language model; on the other hand, by explicitly constraining the expression boundaries of various features through embedding vectors, the large language model can still fully utilize its reasoning and generalization capabilities without being aware of the underlying load mechanism, thereby improving the overall stability and accuracy of load prediction in scenarios with few or no samples.

[0043] Furthermore, in some embodiments of this application, the step of constructing a multidimensional augmented data matrix from the first load data using the Hankel operator, and then performing dynamic pattern decomposition on the multidimensional augmented data matrix to extract dominant feature values, includes: The first load data is sampled using a sliding window of a set length to construct the multidimensional augmented data matrix that captures local dynamic dependencies; Linear operators are constructed based on the mapping relationship of continuous time steps in the multidimensional augmented data matrix to describe the dynamic evolution of the load. Singular value decomposition and eigenvalue decomposition are performed on the linear operator to obtain the dominant eigenvalues ​​that reflect the load change trend and oscillation characteristics.

[0044] Preferably, in some embodiments of this application, the step of sampling the first load data using a sliding window of a set length to construct the multidimensional augmented data matrix that captures local dynamic dependencies includes: Dynamic Mode Decomposition (DMD) is applied to extract the key features underlying the first load data. For the first load data... First, the Hankel operator is used to transform it into a multidimensional augmented data matrix to fully capture the dynamic characteristics of the load sequence. The multidimensional augmented data matrix constructed using the Hankel operator is represented as follows: ; in, express The One element; The window length when constructing a multi-dimensional augmented data matrix for the Hankel operator determines the number of time-series data points contained in each sliding window, and its function is to capture the local dynamic dependencies of building load time-series data.

[0045] Preferably, in some embodiments of this application, the construction of a linear operator based on the mapping relationship of continuous time steps in the multidimensional augmented data matrix includes: assuming that there is a time dependency between load values ​​in the local time neighborhood, i.e., the already acquired observation values They do not exist in isolation, but are intrinsically related to their preceding and following observations. This relationship can be mathematically modeled and described using a q-order autoregressive model: Observations It can be represented as the front A linear combination of observations plus negligible residuals ,Right now: ; in, It is a weight.

[0046] Based on the above assumptions, there must exist a linear operator. satisfy: ;in .

[0047] Therefore, let's assume that... and Given two consecutive input matrices, the following equation applies: ; From this, we can deduce the following formula: ,in This is the final linear operator.

[0048] Preferably, in some embodiments of this application, the step of performing singular value decomposition and eigenvalue decomposition on the linear operator to obtain the dominant eigenvalues ​​reflecting the load change trend and oscillation characteristics includes: Linear operators This is also a matrix, and the eigenvalues ​​of this matrix can reflect the dynamic characteristics of the load sequence, such as the decay trend and the growth trend. Based on this, by extracting... The core features are constructed using the feature values ​​of the features. The specific extraction process is as follows: First, determine the rank. Singular Value Decomposition (SVD) transforms the matrix Approximately ;in, Containing diagonal elements The largest singular value, and Each contains the corresponding left singular vector and right singular vector. This indicates the conjugate transpose operation.

[0049] Then the matrix Projected to dimensional subspace Up, get ; where the projection matrix of eigenvalues ​​and of The dominant eigenvalues ​​overlap.

[0050] Next, the following formula is used to... Perform eigenvalue decomposition ;in, and Each contains eigenvectors and eigenvalues.

[0051] Finally, select the matrix. The former Each eigenvalue is used as a feature, while also noting that... singular values It reflects the intensity of signal components in the load sequence. The dominant eigenvector is constructed using dominant eigenvalues ​​and singular values. .

[0052] Preferably, in some embodiments of this application, generating the feature embedding vector based on the dominant feature value through linear transformation includes: transforming the dominant feature vector... A linear transformation is performed through a linear layer; the process can be described as follows: ; in, and These are the learnable parameters in the feature extraction layer. It is the generated feature embedding vector.

[0053] This application introduces the Hankel operator and dynamic mode decomposition to perform multidimensional augmentation and dominant feature extraction on the load sequence, transforming the originally highly non-stationary and noisy time series into a dynamic feature space that can be characterized by a small number of dominant modes. This approach can effectively separate the long-term trend components and periodic oscillation components in load changes, reduce the interference of short-term random fluctuations on feature representation, and make the extracted feature embedding vectors closer to the intrinsic physical mechanism of load evolution, thus maintaining strong interpretability and predictive robustness even under conditions of insufficient samples.

[0054] Furthermore, in some embodiments of this application, the step of normalizing and aggregating the first load data and the dynamically sampled second load data to obtain aggregated feature data includes: Based on a preset similarity measurement criterion, a set number of first buildings are randomly selected from the adjacent buildings of the same type. The first load data and the second load data corresponding to each of the first buildings are normalized to obtain a standard load sequence. The aggregated characteristic data, which characterizes the shared electricity consumption of the community, is obtained by averaging all the standard load sequences.

[0055] Preferably, in some embodiments of this application, the step of obtaining the feature aggregation data includes: a spatial embedding layer designed to collect information from neighboring nodes to improve prediction performance.

[0056] First, identify adjacent buildings of the same type as the target building. There is a strong spatial correlation in load consumption patterns among buildings of the same type; therefore, it is necessary to select buildings of the same type... Randomly selected from buildings Building ( To construct the target building Adjacent set This application does not select all other buildings as adjacent buildings, as this could lead to data redundancy and degrade model performance. Note that the set of buildings of the same type is constructed through random selection, both during the training and real-time testing phases.

[0057] Secondly, regarding the first load data and its corresponding second load data Perform normalization. Let This represents the normalized second load data. , This represents the first load data after normalization. , and All are standard load sequences. Then, information from the first load data and its corresponding second load data is aggregated in the following manner to obtain aggregated feature data. : ; Since neighbor information aggregation is performed node-by-node (building) aggregation, the number of buildings in the offline training phase may differ from the number in the online testing phase. It is important to note that the number of neighbor buildings differs between the offline training phase and the online testing phase. There are also different options. During the online testing phase, regardless of whether the number of buildings increases or decreases, neighbors can be dynamically sampled and training parameters can be reused to generate features without adjusting the network structure or retraining.

[0058] Obtain aggregated feature data Then, we use the block division method mentioned above to divide it into blocks, obtaining... Then, it is fed into a linear layer, as follows: ; in, and These are learnable parameters in the spatial embedding layer. It is the generated spatial embedding vector.

[0059] This application utilizes dynamic sampling from adjacent buildings of the same type based on similarity metrics, and performs dimensional normalization and mean aggregation on the load sequences of multiple buildings to construct aggregated feature data characterizing the common electricity consumption behavior of a cluster. This effectively mitigates the uncertainty caused by the scarcity or abnormal fluctuations of load data for a single target building. This approach avoids the noise accumulation problem caused by fixing adjacency relationships or fully introducing external building data, enabling the spatial embedding vector to stably reflect the relative position and collaborative characteristics of the target building within the same building group, thereby improving the reliability of the model in cross-building generalization prediction.

[0060] S103: Input the input embedding vector, the spatial embedding vector, and the feature embedding vector into a preset large language model to obtain a load prediction embedding vector representing the load prediction result; wherein, the large language model avoids access to the internal parameters of the large language model by iteratively training the black-box adaptation network.

[0061] Preferably, in some embodiments of this application, reference is made to Figure 2 When the input embedding vector is obtained The spatial embedding vector and the feature embedding vector Afterwards, , , Concatenate the lines into the total input embedding. The pre-trained LLM is input to obtain the load prediction embedding vector. Then, a linear transformation is performed through the output projection layer. Generate target building's predicted load data for the next F steps. .

[0062] The input to a pre-trained language model (LLM) is the connection of all three layers of embeddings, i.e. .

[0063] Furthermore, in some embodiments of this application, the iterative training of the black-box adaptation network using a random approximation algorithm with the zero-order gradient includes: Obtain the first parameter of the black-box adaptation network in the current iteration, generate a random perturbation vector using a preset Bernoulli distribution, and apply the random perturbation vector to the first parameter to obtain a pair of positive perturbation parameters and negative perturbation parameters; The positive perturbation parameters and the negative perturbation parameters are respectively input into the black box adaptation network, and the corresponding positive perturbation loss value and negative perturbation loss value are calculated by combining the inference results of the large language model. Based on the difference between the positive perturbation loss value and the negative perturbation loss value, and in conjunction with the random perturbation vector, the zeroth-order gradient characterizing the evolution direction of the first parameter is calculated. The black-box adaptation network is updated using the zero-order gradient obtained in each iteration, so that the black-box adaptation network tends to converge while avoiding access to the internal parameters of the large language model, thus obtaining the trained black-box adaptation network.

[0064] Preferably, in some embodiments of this application, a training dataset is first obtained before training the black-box adaptation network and subsequent input layers. Assume the historical load dataset corresponding to the target building is... , The total number of buildings in the area is N, that is Target building From time Time The first load data is , express The first in Item data. Let's assume... Indicates the target building A collection of adjacent buildings of the same type, and Indicates from time The second load data for adjacent buildings of the same type up to time t. Let... and They represent from time arrive target building The actual load data and the predicted load data. The prediction formula during the training phase is: .

[0065] Furthermore, the purpose of training the black-box adaptation network and subsequent input layers is to optimize the learnable parameters. This makes the predicted load data closer to the actual load data, that is: ; in, It is the total number of historical time steps; and Representing the first Forecast load data and actual load data at each time step.

[0066] Preferably, in some embodiments of this application, the calculation step of the zeroth-order gradient includes: We employ the classic zero-order gradient estimator—Synchronous Perturbation Stochastic Approximation (SPSA)—to achieve zero-order gradient estimation solely through model inference results. Theoretical studies have confirmed the convergence of SPSA and its superior performance in neural network optimization problems. Its estimation process can be described as follows: ; in, That is, the trainable parameters in the input embedding layer, spatial embedding layer, and feature extraction layer. express The zeroth gradient. From standard Gaussian distribution A vector randomly selected from the data. Size and same. It is the perturbation scale parameter. This is the loss function, which serves to provide a comparison of the errors caused by positive and negative perturbations. This represents the positive disturbance loss value. This represents the loss value due to positive disturbance.

[0067] This application employs a zero-order gradient approximation algorithm based on random perturbations to train a black-box adaptation network, achieving effective optimization of overall prediction performance without accessing or backpropagating the internal parameters of a large language model. On one hand, this training method circumvents the engineering limitations of unavailable parameters and high computational costs associated with large language models, significantly reducing system deployment difficulty. On the other hand, by utilizing the difference between positive and negative perturbation losses to estimate the parameter evolution direction, the black-box adaptation network can gradually learn embedding mapping relationships that better align with the inference preferences of the large language model, thereby achieving efficient collaboration with the large language model.

[0068] S104: Determine the load prediction result of the target building based on the load prediction embedding vector.

[0069] Furthermore, in some embodiments of this application, determining the load forecast result of the target building based on the load forecast embedding vector includes: The load prediction embedding vector is input into the pre-trained input layer, and the load prediction embedding vector is linearly transformed by the input layer to obtain the load prediction result; wherein, the input layer completes iterative training based on the first-order gradient of backpropagation during the pre-training stage.

[0070] Preferably, in some embodiments of this application, the step of inputting the load prediction embedding vector into a pre-trained input layer, and performing a linear transformation on the load prediction embedding vector through the input layer to obtain the load prediction result includes: make This represents the load prediction embedding vector output by the pre-trained Large Language Model (LLM), which is then passed through an output projection layer consisting of a linear input layer. The formula for converting this into load forecasting results is as follows: ; in, and These are the learnable parameters in the input layer. This corresponds to the first load data. The output layer transforms the large language model's hidden space into the load prediction space, completing the final mapping step from feature extraction to the generation of specific predicted values. The output layer's parameters are designed to be shared across buildings, learning a universal mapping relationship between LLM output embeddings and load prediction values. When the number of buildings changes, the embedding vectors of newly added buildings can be directly predicted using these shared parameters, achieving effective results without separate training and enhancing adaptability to dynamic changes in the number of buildings.

[0071] By setting up an independently pre-trained output mapping layer, the load forecast embedding vector is linearly mapped to the final load forecast result, effectively decoupling the semantic reasoning space from the numerical prediction space. This structure avoids the instability caused by directly relying on large language models to generate numerical results, allowing the numerical accuracy of the prediction results to be controlled and optimized through traditional supervised learning methods. Thus, while maintaining the generalization advantages of large language models, it ensures that the load forecast results meet the engineering requirements of power business for accuracy and controllability.

[0072] Furthermore, in some embodiments of this application, the output layer completes iterative training based on the first-order gradient of backpropagation during the pre-training phase, including: The first load prediction result output by the output layer in the current iteration step is compared with the corresponding first real load data, and the prediction loss representing the prediction deviation is calculated. Based on the chain rule, the predicted loss is backpropagated along the output layer to calculate the first gradient at the current iteration step. The output layer is updated based on the first-order gradient obtained in each iteration by using a preset adaptive moment estimation optimizer to minimize the prediction loss.

[0073] Preferably, in some embodiments of this application, the formula for calculating the first-order gradient is: Applying the chain rule to calculate parameters in the input layer and The first-order gradient is as follows: ; in, and They represent and The first gradient.

[0074] This application pre-trains the output layer using first-order gradient backpropagation and an adaptive moment estimation optimizer, enabling efficient and stable minimization of prediction errors. On one hand, the adaptive optimization strategy automatically adjusts the learning rate according to different time scales and load amplitude changes, improving training convergence speed. On the other hand, the independent optimization of the output layer allows the system to adapt to new scenarios by only fine-tuning the mapping layer when facing different buildings and load levels, thereby further enhancing the practicality and scalability of the overall method in applications with few samples and across different scenarios.

[0075] Preferably, in some embodiments of this application, when the zeroth-order gradient and the first-order gradient are obtained, and the corresponding parameters are optimized based on the gradient values, the optimization includes: using Adam to optimize the parameters. (i.e., the trainable parameters in the black-box adaptation network and the output layer), specifically: make Represents all parameters The gradient is calculated. In the first... In this iteration, the parameter optimization process can be described as follows: ; in, and Iterations Partial first-order moment estimator and second-order primitive moment estimator; and For iteration First-order moment estimates and second-order primitive moment estimates after bias correction. To control the hyperparameter of the exponential decay rate, For learning rate, To ensure numerical stability, small constants, These are the trainable parameters updated after iteration t. Termination occurs when the loss function converges. Updates and iterations.

[0076] In summary, the building load prediction method provided in this application has the following advantages compared to existing technologies: By simultaneously introducing the load data of the target building itself and the load data of adjacent buildings of the same type, it explicitly supplements the effective information sources of the target building under the condition of insufficient samples at the data level; and by using a black-box adaptation network to uniformly model the above-mentioned multi-source load data, it extracts multi-dimensional spatiotemporal features that take into account the temporal evolution law, spatial synergy effect, and load dynamic mode, thereby enhancing the ability to characterize the intrinsic mechanism of building load at the feature level. On this basis, the multi-dimensional spatiotemporal features are mapped to the input space of a large language model, enabling the large language model to exert its generalization advantage in few-sample and zero-sample inference without directly perceiving the original load data and spatial structure. At the same time, iterative training through a black-box adaptation network avoids the access restrictions on the internal parameters of the large language model, significantly reducing the difficulty of model adaptation and computational overhead. Finally, the prediction result is mapped through an independently trained output layer, so that the overall method effectively improves the accuracy of building load prediction in few-sample or zero-sample scenarios while ensuring feasibility.

[0077] like Figure 3 As shown, based on the above-mentioned method embodiments, an embodiment of this application provides a building load prediction device, including: a data acquisition module 201, a feature extraction module 202, a model calling module 203, and a mapping module 204; The acquisition module 201 is used to acquire the first load data of the target building and the second load data of several adjacent buildings of the same type corresponding to the target building. The feature extraction module 202 is used to input the first load data and each of the second load data into the pre-trained black-box adaptation network to obtain an input embedding vector characterizing the electricity consumption time sequence pattern of the target building, a spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and a feature embedding vector characterizing the load dynamic mode of the target building. The model invocation module 203 is used to input the input embedding vector, the spatial embedding vector, and the feature embedding vector into a preset large language model to obtain a load prediction embedding vector representing the load prediction result; wherein, the large language model avoids access to the internal parameters of the large language model by iteratively training the black-box adaptation network. The mapping module 204 is used to determine the load prediction result of the target building based on the load prediction embedding vector.

[0078] Further, in some embodiments of this application, the black-box adaptation network includes: an input embedding layer, a spatial embedding layer, and a feature extraction layer; the feature extraction module 202 includes: a first feature extraction unit, a second feature extraction unit, and a third feature extraction unit; the feature extraction module 202 is used to input the first load data and each of the second load data into the pre-trained black-box adaptation network to obtain, respectively, an input embedding vector characterizing the electricity consumption time-series pattern of the target building, a spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and a feature embedding vector characterizing the load dynamic mode of the target building, including: The first feature extraction unit is used to input the first load data and each of the second load data into the pre-trained black box adaptation network, so as to normalize and linearly transform the first load data through the input embedding layer to generate the input embedding vector. The second feature extraction unit is used to input the first load data and each of the second load data into the pre-trained black box adaptation network, so as to perform normalization and aggregation processing on the first load data and the dynamically sampled second load data through the spatial embedding layer to obtain aggregated feature data, and to linearly transform the aggregated feature data to generate the spatial embedding vector. The third feature extraction unit is used to input the first load data and each of the second load data into the pre-trained black-box adaptation network, so that the first load data is constructed into a multidimensional augmented data matrix by using the Hankel operator through the feature extraction layer, the dominant feature value is extracted by dynamic pattern decomposition of the multidimensional augmented data matrix, and the feature embedding vector is generated by linear transformation based on the dominant feature value.

[0079] Furthermore, in some embodiments of this application, the third feature extraction unit is used to construct a multidimensional augmented data matrix from the first load data using the Hankel operator, and to extract dominant feature values ​​from the multidimensional augmented data matrix through dynamic pattern decomposition, including: The first load data is sampled using a sliding window of a set length to construct the multidimensional augmented data matrix that captures local dynamic dependencies; Linear operators are constructed based on the mapping relationship of continuous time steps in the multidimensional augmented data matrix to describe the dynamic evolution of the load. Singular value decomposition and eigenvalue decomposition are performed on the linear operator to obtain the dominant eigenvalues ​​that reflect the load change trend and oscillation characteristics.

[0080] Furthermore, in some embodiments of this application, the second feature extraction unit is used to perform normalization and aggregation processing on the first load data and the dynamically sampled second load data to obtain aggregated feature data, including: Based on a preset similarity measurement criterion, a set number of first buildings are randomly selected from the adjacent buildings of the same type. The first load data and the second load data corresponding to each of the first buildings are normalized to obtain a standard load sequence. The aggregated characteristic data, which characterizes the shared electricity consumption of the community, is obtained by averaging all the standard load sequences.

[0081] Furthermore, in some embodiments of this application, the iterative training of the black-box adaptation network using a random approximation algorithm with the zero-order gradient includes: Obtain the first parameter of the black-box adaptation network in the current iteration, generate a random perturbation vector using a preset Bernoulli distribution, and apply the random perturbation vector to the first parameter to obtain a pair of positive perturbation parameters and negative perturbation parameters; The positive perturbation parameters and the negative perturbation parameters are respectively input into the black box adaptation network, and the corresponding positive perturbation loss value and negative perturbation loss value are calculated by combining the inference results of the large language model. Based on the difference between the positive perturbation loss value and the negative perturbation loss value, and in conjunction with the random perturbation vector, the zeroth-order gradient characterizing the evolution direction of the first parameter is calculated. The black-box adaptation network is updated using the zero-order gradient obtained in each iteration, so that the black-box adaptation network tends to converge while avoiding access to the internal parameters of the large language model, thus obtaining the trained black-box adaptation network.

[0082] Furthermore, in some embodiments of this application, the mapping module 204 is used to determine the load forecast result of the target building based on the load forecast embedding vector, including: The load prediction embedding vector is input into the pre-trained input layer, and the load prediction embedding vector is linearly transformed by the input layer to obtain the load prediction result; wherein, the input layer completes iterative training based on the first-order gradient of backpropagation during the pre-training stage.

[0083] Furthermore, in some embodiments of this application, the output layer completes iterative training based on the first-order gradient of backpropagation during the pre-training phase, including: The first load prediction result output by the output layer in the current iteration step is compared with the corresponding first real load data, and the prediction loss representing the prediction deviation is calculated. Based on the chain rule, the predicted loss is backpropagated along the output layer to calculate the first gradient at the current iteration step. The output layer is updated based on the first-order gradient obtained in each iteration by using a preset adaptive moment estimation optimizer to minimize the prediction loss.

[0084] In summary, the building load prediction device provided in this application has the following advantages compared to the prior art: By simultaneously introducing the load data of the target building itself and the load data of adjacent buildings of the same type, it explicitly supplements the effective information sources of the target building under the condition of insufficient samples at the data level; and by using a black-box adaptation network to uniformly model the above-mentioned multi-source load data, it extracts multi-dimensional spatiotemporal features that take into account the temporal evolution law, spatial synergy effect, and load dynamic mode, thereby enhancing the ability to characterize the intrinsic mechanism of building load at the feature level. On this basis, the multi-dimensional spatiotemporal features are mapped to the input space of a large language model, enabling the large language model to exert its generalization advantage in few-sample and zero-sample reasoning without directly perceiving the original load data and spatial structure. At the same time, iterative training through a black-box adaptation network avoids the access restrictions on the internal parameters of the large language model, significantly reducing the difficulty of model adaptation and computational overhead. Finally, the prediction result is mapped through an independently trained output layer, so that the overall method effectively improves the accuracy of building load prediction in few-sample or zero-sample scenarios while ensuring feasibility.

[0085] It is understood that the above-described device embodiments correspond to the method embodiments of this application, and can implement the building load prediction method provided by any of the above-described method embodiments of this application.

[0086] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided in this application, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0087] Based on the above embodiments of the building load forecasting method, another embodiment of this application provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the building load forecasting method of any embodiment of this application.

[0088] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete this application. The one or more module units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0089] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0090] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0091] Based on the above-described method embodiments, another embodiment of this application provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the building load prediction method described in any of the above-described method embodiments of this application.

[0092] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

Claims

1. A method for predicting building loads, characterized in that, include: Collect the first load data of the target building and the second load data of several adjacent buildings of the same type corresponding to the target building; The first load data and each of the second load data are input into the pre-trained black-box adaptation network to obtain the input embedding vector characterizing the electricity consumption time series pattern of the target building, the spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and the feature embedding vector characterizing the load dynamic mode of the target building. The input embedding vector, the spatial embedding vector, and the feature embedding vector are input into a preset large language model to obtain a load prediction embedding vector representing the load prediction result; wherein, the large language model is trained iteratively on the black-box adaptation network to avoid access to the internal parameters of the large language model; The load forecast result for the target building is determined based on the load forecast embedding vector; The black-box adaptation network includes: an input embedding layer, a spatial embedding layer, and a feature extraction layer; the first load data and each of the second load data are input into the pre-trained black-box adaptation network to obtain an input embedding vector characterizing the electricity consumption time-series pattern of the target building, a spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and a feature embedding vector characterizing the load dynamic mode of the target building, including: The first load data and each of the second load data are input into the pre-trained black-box adaptation network, so that the first load data is normalized and linearly transformed through the input embedding layer to generate the input embedding vector. The first load data and each of the second load data are input into the pre-trained black-box adaptation network. The first load data and the dynamically sampled second load data are normalized and aggregated through the spatial embedding layer to obtain aggregated feature data. The aggregated feature data is then linearly transformed to generate the spatial embedding vector. The first load data and each of the second load data are input into the pre-trained black-box adaptation network. The first load data is constructed into a multidimensional augmented data matrix by using the Hankel operator through the feature extraction layer. The dominant feature value is extracted by dynamic pattern decomposition of the multidimensional augmented data matrix, and the feature embedding vector is generated by linear transformation based on the dominant feature value.

2. The building load forecasting method as described in claim 1, characterized in that, The step of constructing a multidimensional augmented data matrix from the first load data using the Hankel operator, and then performing dynamic pattern decomposition on the multidimensional augmented data matrix to extract dominant feature values, includes: The first load data is sampled using a sliding window of a set length to construct the multidimensional augmented data matrix that captures local dynamic dependencies; Linear operators are constructed based on the mapping relationship of continuous time steps in the multidimensional augmented data matrix to describe the dynamic evolution of the load. Singular value decomposition and eigenvalue decomposition are performed on the linear operator to obtain the dominant eigenvalues ​​that reflect the load change trend and oscillation characteristics.

3. The building load prediction method as described in claim 1, characterized in that, The process of normalizing and aggregating the first load data and the dynamically sampled second load data to obtain aggregated feature data includes: Based on a preset similarity measurement criterion, a set number of first buildings are randomly selected from the adjacent buildings of the same type. The first load data and the second load data corresponding to each of the first buildings are normalized to obtain a standard load sequence. The aggregated characteristic data, which characterizes the shared electricity consumption of the community, is obtained by averaging all the standard load sequences.

4. The building load prediction method as described in claim 1, characterized in that, The iterative training of the black-box adaptation network using the zero-order gradient through a stochastic approximation algorithm includes: Obtain the first parameter of the black-box adaptation network in the current iteration, generate a random perturbation vector using a preset Bernoulli distribution, and apply the random perturbation vector to the first parameter to obtain a pair of positive perturbation parameters and negative perturbation parameters; The positive perturbation parameters and the negative perturbation parameters are respectively input into the black box adaptation network, and the corresponding positive perturbation loss value and negative perturbation loss value are calculated by combining the inference results of the large language model. Based on the difference between the positive perturbation loss value and the negative perturbation loss value, and in conjunction with the random perturbation vector, the zeroth-order gradient characterizing the evolution direction of the first parameter is calculated. The black-box adaptation network is updated using the zero-order gradient obtained in each iteration, so that the black-box adaptation network tends to converge while avoiding access to the internal parameters of the large language model, thus obtaining the trained black-box adaptation network.

5. The building load forecasting method as described in claim 1, characterized in that, The determination of the load forecast result of the target building based on the load forecast embedding vector includes: The load prediction embedding vector is input into the pre-trained input layer, and the load prediction embedding vector is linearly transformed by the input layer to obtain the load prediction result; wherein, the input layer completes iterative training based on the first-order gradient of backpropagation during the pre-training stage.

6. The building load prediction method as described in claim 5, characterized in that, The output layer undergoes iterative training during the pre-training phase based on the first-order gradients from backpropagation, including: The first load prediction result output by the output layer in the current iteration step is compared with the corresponding first real load data, and the prediction loss representing the prediction deviation is calculated. Based on the chain rule, the predicted loss is backpropagated along the output layer to calculate the first gradient at the current iteration step. The output layer is updated based on the first-order gradient obtained in each iteration by using a preset adaptive moment estimation optimizer to minimize the prediction loss.

7. A building load prediction device, characterized in that, include: The module includes a data acquisition module, a feature extraction module, a model invocation module, and a mapping module. The acquisition module is used to acquire the first load data of the target building and the second load data of several adjacent buildings of the same type corresponding to the target building. The feature extraction module is used to input the first load data and each of the second load data into the pre-trained black-box adaptation network to obtain an input embedding vector characterizing the electricity consumption time sequence pattern of the target building, a spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and a feature embedding vector characterizing the load dynamic mode of the target building. The model invocation module is used to input the input embedding vector, the spatial embedding vector, and the feature embedding vector into a preset large language model to obtain a load prediction embedding vector representing the load prediction result; wherein, the large language model avoids access to the internal parameters of the large language model by iteratively training the black-box adaptation network. The mapping module is used to determine the load forecast result of the target building based on the load forecast embedding vector; The black-box adaptation network includes: an input embedding layer, a spatial embedding layer, and a feature extraction layer; the feature extraction module includes: a first feature extraction unit, a second feature extraction unit, and a third feature extraction unit; the feature extraction module is used to input the first load data and each of the second load data into the pre-trained black-box adaptation network to obtain, respectively, an input embedding vector characterizing the electricity consumption time-series pattern of the target building, a spatial embedding vector characterizing the synergistic effect between the target building and the adjacent buildings of the same type, and a feature embedding vector characterizing the load dynamic mode of the target building, including: The first feature extraction unit is used to input the first load data and each of the second load data into the pre-trained black box adaptation network, so as to normalize and linearly transform the first load data through the input embedding layer to generate the input embedding vector. The second feature extraction unit is used to input the first load data and each of the second load data into the pre-trained black box adaptation network, so as to perform normalization and aggregation processing on the first load data and the dynamically sampled second load data through the spatial embedding layer to obtain aggregated feature data, and to linearly transform the aggregated feature data to generate the spatial embedding vector. The third feature extraction unit is used to input the first load data and each of the second load data into the pre-trained black-box adaptation network, so that the first load data is constructed into a multidimensional augmented data matrix by using the Hankel operator through the feature extraction layer, the dominant feature value is extracted by dynamic pattern decomposition of the multidimensional augmented data matrix, and the feature embedding vector is generated by linear transformation based on the dominant feature value.

8. A terminal device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements a building load forecasting method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform a building load forecasting method as described in any one of claims 1 to 6.

Images