Electric power bid inviting and purchasing scale prediction method and system

Abstract

The invention discloses an electric power bid inviting and purchasing scale prediction method and system. The method comprises the following steps: firstly, acquiring historical electric power bid invitation purchase data, including unstructured text data and structured data, and establishing a cross-modal fusion model and a purchase scale prediction model; thirdly, dividing the historical data into a plurality of meta-learning task sets according to a preset electric power bid invitation purchasing scene, and performing meta-training on the model parameters by adopting a meta-learning framework to enable the model to obtain learning ability; in the real-time prediction application, the system monitors a preset condition, and when the condition is triggered, the newly acquired real-time data and the real purchase scale are used as small samples, and the purchase meta-model parameters are quickly updated online. According to the method, deep mining of multi-source heterogeneous data is realized, the adaptability and prediction precision of the model to a new bid inviting scene are enhanced, and powerful technical support is provided for accurate purchase decision making of the power industry.

Description

Technical Field

[0001] This disclosure relates to the fields of power system automation and artificial intelligence technology, and in particular to a method and system for predicting the scale of power bidding and procurement. Background Technology

[0002] Procurement of power equipment is a fundamental step in ensuring the stable operation of the power grid and the construction of power projects. The determination of procurement scale must strictly adhere to technical specifications, project demand plans, and power grid development plans. Accurate procurement scale forecasting plays a crucial role in optimizing inventory costs, ensuring material supply, and improving supply chain efficiency.

[0003] Currently, in the field of power procurement scale forecasting, some schemes exist that utilize multi-dimensional data for prediction. For example, Chinese patent CN116128105A discloses a method for constructing a power material procurement forecasting model, which attempts to integrate multiple dimensions such as historical bidding and procurement data, annual budget data, annual project reserve data, potential supplier data, and objective factor data, aiming to improve the accuracy of the forecast. However, such existing technologies still share the following common technical shortcomings: First, data feature mining is insufficient, making it difficult to utilize unstructured information. Existing forecasting methods mainly rely on structured fields such as project type, material category, and funding scale for modeling. However, unstructured text data such as technical specifications and project requirements statements contained in power bidding documents often contain technical parameters and special construction requirements that have a decisive impact on the procurement scale. Existing technologies lack effective cross-modal feature extraction methods and cannot understand this deep semantic information. This leads to an inherent bias in the understanding of procurement needs, thus limiting forecasting accuracy.

[0004] Furthermore, these models exhibit weak generalization ability and lack the capacity for small-sample adaptation to new scenarios. Existing technical solutions mostly employ traditional statistical models or conventional supervised learning models (such as linear regression and random forests). These models typically assume that future data follows the same probability distribution as historical data, and that model parameters are fixed once trained, or require the accumulation of large amounts of new data for retraining. In real-world scenarios involving frequent adjustments to power grid development plans, macroeconomic policies, or commodity prices, these static models struggle to effectively utilize limited new data for rapid online fine-tuning. This results in severely lagging predictions in new scenarios, easily leading to excessive procurement and stockpiling or supply shortages.

[0005] In summary, due to deficiencies in multimodal semantic understanding and small-sample dynamic adaptation, existing technologies struggle to meet the practical needs of lean management of power resources. Therefore, there is an urgent need in this field for a novel power bidding and procurement scale prediction scheme. This scheme should deeply integrate knowledge from the power sector, fully utilize multi-source data including unstructured text, and possess adaptive learning capabilities to overcome the shortcomings of existing technologies and achieve more accurate and reliable predictions. Summary of the Invention

[0006] To address the shortcomings of existing technologies in fusing unstructured technical requirement texts with structured power grid operation environment data, and the difficulty in rapidly adapting to constantly changing power bidding scenarios, this invention provides a method and system for predicting the scale of power bidding and procurement. The method first acquires historical power bidding and procurement data, including unstructured text data and structured data, and establishes a cross-modal fusion model and a procurement scale prediction model. Next, the historical data is divided into multiple meta-learning task sets according to preset power bidding and procurement scenarios. A meta-learning framework is used to meta-train the model parameters, enabling the model to acquire learning capabilities. In real-time prediction applications, the system monitors preset conditions and, when these conditions are triggered, uses newly acquired real-time data and actual procurement scales as small samples to rapidly update the procurement meta-model parameters online. This invention achieves deep mining of multi-source heterogeneous data, enhances the model's adaptability to new bidding scenarios and its prediction accuracy, and provides strong technical support for precise procurement decisions in the power industry.

[0007] The present invention adopts the following technical solution.

[0008] Firstly, a method for predicting the scale of power bidding and procurement includes: Step 1: Obtain historical power bidding and procurement data, including unstructured text data representing technical requirements and structured data representing the business operating environment; Step 2: Establish a cross-modal fusion model to perform fusion processing on the unstructured text data and the structured data to generate scale prediction fusion features; and establish a procurement scale prediction model to perform prediction processing on the scale prediction fusion features to obtain the power bidding procurement scale. Step 3: Divide the historical power bidding and procurement data into multiple meta-learning task sets according to the preset power bidding and procurement scenario; adopt the meta-learning framework, and perform meta-training on the cross-modal fusion model parameters and procurement scale prediction model parameters based on the multiple meta-learning task sets to obtain the updated cross-modal fusion model parameters and procurement scale prediction model parameters. Step 4: Obtain the power bidding and procurement data to be predicted, and perform fusion processing and prediction processing based on the updated cross-modal fusion model parameters and procurement scale prediction model parameters to obtain the power bidding and procurement scale. Step 5: Monitor whether the completed power bidding and procurement data has triggered preset conditions; if the preset conditions are triggered, use the completed power bidding and procurement data and the corresponding actual procurement scale as a small sample to update the procurement scale prediction model parameters.

[0009] Preferably, the historical power bidding and procurement data includes: The unstructured text data representing technical requirements includes technical requirement documents for power projects; The structured data representing the business operation environment includes categorical data and numerical data; among which, categorical data includes project attributes, power grid load forecast levels, and seasonal characteristic markers; numerical data includes commodity price indices. In addition, historical real procurement scale data corresponding to the unstructured text data and the structured data.

[0010] Preferably, in step 2, the cross-modal fusion model includes: A text encoder is used to perform word segmentation on the unstructured text data to generate a token sequence; the token sequence is then input into a pre-trained language model to generate a technology requirement embedding vector. A structured data encoder is used to perform entity embedding processing on categorical data in the structured data to generate categorical feature vectors; the categorical feature vectors are concatenated with numerical data in the structured data and input into a multilayer perceptron for nonlinear transformation to generate an operational environment embedding vector. A cross-modal fusion processor is used to receive and fuse the technology requirement embedding vector and the operating environment embedding vector to obtain the scale prediction fusion feature.

[0011] Preferably, the structure of the cross-modal fusion processor includes: The concatenation and embedding layer is used to receive the technical requirement embedding vector and the operating environment embedding vector; concatenate the technical requirement embedding vector and the operating environment embedding vector in the sequence dimension to form a unified vector sequence, and add position encoding and modality type encoding to the unified vector sequence; The system comprises an N-layer Transformer encoder layer for receiving the output of the splicing and embedding layer, and each layer includes a multi-head self-attention sublayer and a feedforward network sublayer. The multi-head self-attention sublayer is used to perform self-attention computation on the unified vector sequence, and the feedforward network sublayer is used to perform nonlinear transformation on the output of the multi-head self-attention sublayer. The feature pooling layer is used to receive the final unified vector sequence output by the last layer in the N-layer Transformer encoder layer; and to perform a pooling operation on the final unified vector sequence to generate and output the scale prediction fusion feature.

[0012] Preferably, in step 2, the structure of the procurement scale prediction model includes: An input layer is used to receive the scale prediction fusion features output by the cross-modal fusion unit; A hidden layer with a nonlinear activation function is used to perform a nonlinear transformation on the scale prediction fusion features; And an output layer with linearly activated neurons to obtain the scale of the power bidding and procurement.

[0013] Preferably, in step 3, a meta-learning framework is used to perform meta-training on the parameters of the cross-modal fusion model and the parameters of the procurement scale prediction model based on multiple meta-learning task sets, including: Each meta-learning task in the multiple meta-learning task sets is divided into a support set and a query set; M meta-learning tasks are sampled from the set of multiple meta-learning tasks for iterative training, the iterative training including an inner loop stage and an outer loop stage; In the inner loop phase, for each of the M meta-learning tasks, the current procurement scale prediction model parameters are loaded, the loss is calculated using the data of the corresponding support set, and gradient descent is performed to update the procurement scale prediction model parameters to obtain task-specific parameters. In the outer loop phase, the prediction loss is calculated on the corresponding query set using the task-specific parameters and the current cross-modal fusion model parameters, and the prediction losses of the M meta-learning tasks are summed to obtain the meta-loss. Based on the meta-loss, gradient descent is simultaneously performed on the parameters of the cross-modal fusion model and the parameters of the procurement scale prediction model to obtain the updated parameters of the cross-modal fusion model and the parameters of the procurement scale prediction model.

[0014] Preferably, in step 5, the completed power bidding and procurement data and the corresponding actual procurement scale are used as a small sample to update the parameters of the procurement scale prediction model, including: Load the cross-modal fusion model parameters and the procurement scale prediction model parameters obtained by the meta-training; The cross-modal fusion model parameters are used to perform fusion processing on the data in the small sample to generate scale prediction fusion features; The parameters of the procurement scale prediction model are used to perform prediction processing on the scale prediction fusion features, and the prediction results are compared with the actual procurement scale in the small sample to calculate the small sample prediction loss. Based on the small sample prediction loss, gradient descent is performed to update the parameters of the procurement scale prediction model to obtain the updated procurement scale prediction model parameters.

[0015] Secondly, a power bidding and procurement scale prediction system, which operates the aforementioned power bidding and procurement scale prediction method, includes: The data acquisition and preparation module is used to acquire historical power bidding and procurement data, including unstructured text data representing technical requirements and structured data representing the business operation environment. A cross-modal model building module is used to establish a cross-modal fusion model, which performs fusion processing on the unstructured text data and the structured data to generate scale prediction fusion features; and to establish a procurement scale prediction model, which performs prediction processing on the scale prediction fusion features to obtain the power bidding procurement scale. The meta-training and parameter generation module is used to divide the historical power bidding and procurement data into multiple meta-learning task sets according to the preset power bidding and procurement scenario; using a meta-learning framework, based on the multiple meta-learning task sets, meta-training is performed on the cross-modal fusion model parameters and procurement scale prediction model parameters to obtain updated cross-modal fusion model parameters and procurement scale prediction model parameters. The real-time forecasting module is used to acquire the power bidding and procurement data to be predicted, and based on the loaded and updated cross-modal fusion model parameters and procurement scale forecasting model parameters, it performs fusion processing and forecasting processing to obtain the power bidding and procurement scale. The online monitoring and adaptive update module is used to monitor whether the completed power bidding and procurement data has triggered preset conditions; if the preset conditions are triggered, the completed power bidding and procurement data and the corresponding actual procurement scale are used as a small sample to update the parameters of the procurement scale prediction model.

[0016] Thirdly, a terminal includes a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method.

[0017] Fourthly, a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method.

[0018] The beneficial effects of this invention are as follows: 1. This invention establishes a cross-modal fusion model that deeply integrates unstructured textual data representing technical requirements (such as technical specifications) with structured data representing the business operation environment (such as power grid load levels and project attributes). This overcomes the shortcomings of traditional methods that rely solely on historical time-series data or struggle to align multimodal data. It effectively captures the complex nonlinear dependencies between specific technical requirements and specific operating environments, thereby making the prediction model closer to the actual business decision-making logic.

[0019] 2. This invention does not train a static predictive model. Instead, it divides historical data into multiple meta-learning task sets according to procurement scenarios unique to the power industry, such as peak summer demand, new construction projects, and high load levels. Through meta-training using a meta-learning framework, the model learns how to learn, that is, it pre-masters the meta-knowledge and optimal parameters for quickly adapting from one power scenario to another.

[0020] 3. Based on the model parameters obtained from meta-learning training, this invention can update the model online by using only a very small number of new samples (small samples) and performing one or more gradient descent operations only on the parameters of the procurement scale prediction model when a new situation or performance degradation is detected. This mechanism replicates the parameter optimization logic of the inner loop in the meta-training stage, effectively alleviating the forgetting problem faced by traditional deep learning models during continuous learning, and eliminating the need for time-consuming model retraining, thus achieving high timeliness, high accuracy, and low maintenance costs simultaneously. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for predicting the scale of power bidding and procurement provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0023] Example 1: This embodiment is based on the historical bidding and procurement data of power companies to illustrate the implementation process of the method of the present invention more clearly.

[0024] like Figure 1 As shown, the present invention provides a method for predicting the scale of power bidding and procurement, including: Step 1: Obtain historical power bidding and procurement data, including unstructured text data representing technical requirements and structured data representing the business operating environment.

[0025] The historical power bidding and procurement data includes: The unstructured text data representing technical requirements includes technical requirements documents for power projects. These documents can be in PDF, Word, or text format. After acquiring the data, the system will extract key text content using text parsing or optical character recognition (OCR) technology. The content includes at least technical details such as project type, voltage level, equipment capacity, construction period, and equipment list.

[0026] The structured data representing the business operation environment includes categorical data and numerical data. This data can be obtained from the power company's internal dispatch and forecasting system, project management database, or public economic databases (such as Wind and data from the National Bureau of Statistics), and is time-series aligned based on the project's timestamp.

[0027] The data is categorized into three types: project attributes, power grid load forecast levels, and seasonality indicators. Numerical data includes commodity price indices. Project attributes distinguish the fundamental type of project, such as being specifically coded as "newly constructed power grid project" (corresponding to large-scale infrastructure) or "power grid operation and maintenance upgrade" (corresponding to equipment replacement). Power grid load forecast levels characterize the short-term operational pressure on the power grid, such as being specifically coded as "high," "medium," and "low" (one-hot coding or ordinal coding). Seasonality indicators characterize the seasonal operational status of the power grid, such as being specifically coded as "peak summer demand," "peak winter demand," and "stable period," reflecting the varying urgency of material demand in different seasons. Commodity price indices characterize the transmission effect of upstream raw material costs on procurement scale, such as using monthly or quarterly average price indices for copper, aluminum, or steel.

[0028] In addition, historical real procurement scale data corresponding to the unstructured text data and the structured data. The procurement scale data (e.g., the final winning bid amount or the quantity of materials procured) serves as the target output of the supervised learning model and is used for loss calculation in the subsequent model training phase.

[0029] Step 2: Based on the historical power bidding and procurement data obtained in Step 1, establish a cross-modal fusion model to perform fusion processing on the unstructured text data and the structured data to generate scale prediction fusion features; and establish a procurement scale prediction model to perform prediction processing on the scale prediction fusion features to obtain the power bidding and procurement scale.

[0030] Step 2.1 Establish a cross-modal fusion model The cross-modal fusion model mainly consists of a text encoder, a structured data encoder, and a cross-modal fusion unit, including: A text encoder is used to perform word segmentation on the unstructured text data to generate a token sequence; the token sequence is then input into a pre-trained language model to generate a technology requirement embedding vector.

[0031] Specifically, firstly, the input power project technical requirements document is subjected to data parsing and cleaning, including: converting PDF, Word and other file formats into a processable plain text format; for unstructured tables embedded in the text (such as equipment list tables, technical parameter tables, etc.), key-value pair serialization technology is used to convert them into text descriptions and insert them into the original text; the text is denoised, and irrelevant characters such as headers, footers, numbers, and invalid symbols are deleted.

[0032] Secondly, word segmentation and domain terminology standardization are performed. A sub-word-based segmentation algorithm (such as WordPiece) is used to convert the text into a token sequence; for power industry terms (such as main transformer capacity, primary equipment, GIS switchgear, etc.), a domain dictionary or word vector extension model is used for standardization and synonym mapping to maintain semantic consistency of the same technical meanings across different documents.

[0033] Then, the token sequence is input into a pre-trained language model to generate a technology requirement embedding vector. The pre-trained language model can be a Transformer-based contextual representation model, such as BERT or RoBERTa fine-tuned with an electrical corpus. The model calculates the inter-word contextual relationships through a multi-layer self-attention mechanism and outputs a contextualized vector for each token; the [CLS] vector or the pooled overall text vector is used as the global embedding vector for the project's technology requirements.

[0034] Optionally, to enhance the model's ability to express the semantics of power technology, secondary domain adaptive training (such as using a masked language modeling task) can be performed on the basis of the original pre-trained model using a corpus of power project technical documents, and supervised fine-tuning can be performed by combining historical procurement scale label data. The output of this text encoder is the technical requirement embedding vector, which provides the technical semantic representation input for the subsequent cross-modal fusion layer.

[0035] A structured data encoder is used to perform entity embedding processing on categorical data in the structured data to generate categorical feature vectors; the categorical feature vectors are concatenated with numerical data in the structured data and input into a multilayer perceptron for nonlinear transformation to generate an operational environment embedding vector.

[0036] First, the input structured data is subjected to format normalization and missing value handling. For categorical data, one-hot encoding or a trainable embedding matrix is ​​used to map discrete categories into low-dimensional continuous vector representations; for numerical data, normalization or standardization is performed to eliminate dimensional differences.

[0037] Then, the categorical feature vector and the numerical feature vector are concatenated to form a unified structured input feature vector, which is then input into a multi-layer perceptron (MLP) to perform nonlinear mapping. The MLP includes several hidden layers, each containing nonlinear activation units (such as ReLU, GELU, or Tanh) and dropout layers to prevent overfitting and improve feature representation capabilities.

[0038] Optionally, to enhance the model's ability to perceive temporal and environmental fluctuations, a time-series modeling unit can be introduced into the structured data encoder. For example, when indicators such as commodity price indices and power grid load forecast levels have time-series characteristics, a one-dimensional convolutional layer (1D-CNN) or a gated recurrent unit (GRU) can be used to model their temporal dynamics, thereby extracting short-term trend features. The short-term trend features are then concatenated and fused with the static structured features extracted by the multilayer perceptron to generate the final operating environment embedding vector.

[0039] This operational environment embedding vector can comprehensively reflect the macroeconomic status, power demand level, and seasonal dispatch characteristics of the project, providing structured semantic input for the cross-modal fusion layer.

[0040] A cross-modal fusion processor is used to receive and fuse the technology requirement embedding vector and the operating environment embedding vector, and output the scale prediction fusion feature.

[0041] Since the embedded vectors for technical requirements and the embedded vectors for the operational environment may differ in semantic space and dimension, the cross-modal fusion builder projects the two types of vectors into a unified latent representation space through an independent linear mapping layer in order to perform subsequent intermodal fusion calculations.

[0042] The structure of the cross-modal fusion device includes: The concatenation and embedding layer receives the technology requirement embedding vector and the operating environment embedding vector. These are projected onto the same feature dimension through two independent linear layers. The technology requirement embedding vector and the operating environment embedding vector are concatenated along the sequence dimension to form a unified vector sequence. Position encoding (to identify the relative order of vectors in the sequence) and modality type encoding (such as "text modality encoding" or "structured modality encoding") are added to the unified vector sequence to explicitly identify its source and relative order, thereby enabling the model to distinguish features from different sources.

[0043] An N-layer Transformer encoder layer is used to receive the output of the splicing and embedding layer, and each layer includes a multi-head self-attention sublayer and a feedforward network sublayer. The multi-head self-attention sublayer is used to perform self-attention calculation on the unified vector sequence to dynamically capture the correlation between technical demand information and operating environment information (e.g., learning the nonlinear coupling relationship between "high voltage level" technical demand and "copper price increase" environmental factors). The feedforward network sublayer is used to perform nonlinear transformation on the output of the multi-head self-attention sublayer to improve the high-order feature representation capability.

[0044] Optionally, to further enhance the interaction capabilities between modalities, modal gating weights can be introduced into the multi-head attention mechanism. This involves assigning learnable weight coefficients to different modalities to control the contribution ratio of each modality in feature fusion, thereby enabling the model to adaptively adjust the fusion strategy according to different project features.

[0045] The feature pooling layer is used to receive the final unified vector sequence output by the last layer in the N-layer Transformer encoder layer; and to perform a pooling operation on the final unified vector sequence, which can extract the vector at the first position of the sequence or perform average pooling on the entire sequence vector to generate and output the scale prediction fusion feature.

[0046] Through the aforementioned cross-modal fusion engine, the model can achieve deep semantic alignment and dynamic weight fusion of textual and structured modalities, thereby improving the generalization ability and accuracy of procurement scale prediction.

[0047] Step 2.2 Establish a procurement scale forecasting model This model serves as the downstream task head, used to output the final numerical value based on the fused features. The structure of the procurement scale prediction model includes: An input layer is used to receive the scale prediction fusion features output by the cross-modal fusion unit. These features contain joint information about technical requirements and the operating environment.

[0048] The hidden layer consists of multiple fully connected layers (FCL), each with a non-linear activation function to perform a non-linear transformation on the scale prediction fusion features. Dropout layers are added between each hidden layer to prevent overfitting and ensure that the model has good generalization ability.

[0049] And an output layer, with linearly activated neurons, to output the scale of the electricity tender. This output represents the predicted quantity of materials to be procured or the budget size.

[0050] The procurement scale prediction model is trained using historical electricity procurement data, employs cross-validation to evaluate its generalization performance, and utilizes an early stopping mechanism to prevent overfitting. After training, the model can combine historical data with real-time environmental changes to provide high-precision prediction results for electricity material procurement.

[0051] Based on the two structures constructed in step 2, in step 3, the historical power bidding and procurement data is divided into multiple meta-learning task sets according to the preset power bidding and procurement scenario; using a meta-learning framework, the cross-modal fusion model parameters and procurement scale prediction model parameters are meta-trained based on the multiple meta-learning task sets to obtain updated cross-modal fusion model parameters and procurement scale prediction model parameters.

[0052] Step 3.1 Construction of the Meta-Learning Task Set The preset power bidding and procurement scenario is divided based on the structured data to obtain the multiple meta-learning task sets.

[0053] Specifically, the partitioning process includes: The structured data fields are extracted from historical power bidding and procurement data, including project attribute fields, power grid load forecast level fields, and seasonal characteristic marker fields. Based on the value characteristics of the field, the sample data can be grouped using an automatic clustering method based on preset rules or clustering algorithms (such as K-means) to form multiple different power bidding and procurement scenarios.

[0054] In this embodiment, the plurality of meta-learning task sets include one or more of the following: The task set is divided according to the attributes of the project, and the task set includes procurement data tasks for new power grid projects and procurement data tasks for power grid operation and maintenance renovation. The task set is divided according to the power grid load forecast level, and the task set includes procurement data tasks under the high load forecast level and procurement data tasks under the low load forecast level. The task set is divided according to the seasonal characteristic markers, and the task set includes procurement data tasks during the peak summer cycle and procurement data tasks during the peak winter cycle.

[0055] Using the above method, historical sample data can be automatically divided into different power procurement scenarios based on the values ​​of structured data fields. Each task set represents a typical business scenario for subsequent meta-learning training.

[0056] Step 3.2 Meta-training Iteration Process In step 3, a meta-learning framework is adopted, and meta-training is performed on the parameters of the cross-modal fusion model and the parameters of the procurement scale prediction model based on the multiple meta-learning task sets, including: First, each meta-learning task in the multiple meta-learning task sets is divided into a support set and a query set. Each meta-learning task set... This represents a single electricity procurement scenario, such as "high-load summer procurement" or "low-load winter operation and maintenance upgrades." Each meta-learning task set contains several sample data points. , denoted as:

[0057] in, This represents the original input data for the j-th sample. Labels indicating the scale of purchases. This represents the total number of historical samples in the i-th meta-learning task set. For each meta-learning task set... Further divided into support sets and query set The support set is used for in-task learning of the model, while the query set is used for model evaluation and meta-gradient updates. The split ratio can be set according to the sample size, for example, 8:2.

[0058] Secondly, from the multiple meta-learning task sets M meta-learning tasks are sampled and iteratively trained, and the iterative training includes an inner loop stage and an outer loop stage; where K represents the number of meta-learning task sets.

[0059] In the inner loop phase, for each of the M meta-learning tasks... Load the current procurement scale prediction model parameters Maintain cross-modal fusion model parameters The parameters remain unchanged. The loss is calculated using samples from the corresponding support set, and gradient descent is performed to update the parameters of the procurement scale prediction model to obtain task-specific parameters, specifically including: For each support set sample First, the scale prediction fusion features are calculated using a cross-modal fusion model. :

[0060] in, It is a cross-modal fusion model, using parameters Processing raw input Output scale prediction fusion features .

[0061] Then, the scale prediction fusion features are input into the procurement scale prediction model:

[0062] in, It is a purchasing scale prediction model (such as MLP), using parameters deal with Output the predicted procurement scale .

[0063] Calculate the prediction loss function For example, using the mean squared error (MSE) form:

[0064] Based on this loss, one or more gradient descent updates are performed on the parameters of the procurement scale prediction model to obtain task-specific parameters. :

[0065] in, For the inner loop learning rate, Indicates the calculation of the loss function Compared to The gradient.

[0066] In the outer loop phase, the task-specific parameters and the current cross-modal fusion model parameters are used in the corresponding query set. The prediction loss is calculated by summing the prediction losses of the M meta-learning tasks to obtain the meta-loss, which specifically includes: For each Keep updated task-specific parameters And the current cross-modal fusion model parameters Calculate the query set prediction loss :

[0067] in, .

[0068] The meta-loss is obtained by summing or weighting the query set losses of all tasks. :

[0069] Based on the meta-loss, gradient descent is simultaneously performed to update the parameters of the cross-modal fusion model and the parameters of the procurement scale prediction model, resulting in updated parameters of the cross-modal fusion model and the procurement scale prediction model:

[0070]

[0071] in, The learning rate is the outer loop rate.

[0072] After multiple rounds of iterative training, when the meta-loss When the convergence occurs or the decrease rate falls below a preset threshold for several consecutive rounds, the update process stops and the optimized cross-modal fusion model parameters are output. And procurement scale prediction model parameters .

[0073] By introducing a meta-learning framework, the model learns generalized parameter update directions on the support set of multiple tasks and performs meta-level optimization on the query set, thereby improving the model's generalization and transfer capabilities in different power bidding and procurement scenarios and reducing retraining costs.

[0074] Step 4: Obtain the power bidding and procurement data to be predicted. The format of the power bidding and procurement data is consistent with the format of the historical power bidding and procurement data in Step 1, including real-time unstructured text data and real-time structured data. Call the model with loaded parameters, set the model to inference state, and disable the random dropout function of the Dropout layer described in Step 2 to ensure the determinism of the prediction results. Following the same data processing flow described in Step 2, and based on the loaded and updated cross-modal fusion model parameters and procurement scale prediction model parameters, perform fusion processing and prediction processing, and output the predicted power bidding and procurement scale. Through this step, real-time prediction of the power bidding and procurement scale can be achieved, ensuring that the model can provide dynamic response capabilities in a continuously changing power market environment.

[0075] Step 5: Monitor whether the completed power bidding and procurement data has triggered preset conditions; if the preset conditions are triggered, use the completed power bidding and procurement data and the corresponding actual procurement scale as a small sample to update the procurement scale prediction model parameters.

[0076] Because the number of parameters in the procurement scale prediction model (prediction head) is much smaller than that in the cross-modal fusion model (backbone), the computational cost of updating only the prediction head is extremely low. This allows the model to immediately perform adaptive updates when new data (small samples) arrives, meeting the needs of real-time prediction. Furthermore, the backbone stores general fusion knowledge learned from all historical data during meta-training. If only small samples are used to update the backbone here, the model would suffer from severe overfitting and forget the historical general knowledge.

[0077] In this embodiment, the preset condition is triggered, which may include one or more of the following situations: Performance monitoring-based triggering: When the cumulative error of the model in real-time prediction (e.g., root mean square error RMSE or mean absolute percentage error MAPE) exceeds a preset performance threshold, it indicates that the model's adaptability to the current data has decreased, triggering an update.

[0078] Triggered by environmental changes: An update is triggered when a pre-defined significant change in the business operation environment is detected in the real-time power bidding and procurement data. For example, the internal dispatch system issues an "unexpected high temperature warning," or a new "engineering project attribute" appears in the project management database (e.g., a completely new type of project is approved).

[0079] When the preset conditions are triggered, in step 5, the completed power bidding and procurement data and the corresponding actual procurement scale are used as a small sample to update the parameters of the procurement scale prediction model, including: Obtain recently completed power bidding and procurement data and the corresponding actual procurement scale to construct a small sample set.

[0080] The cross-modal fusion model obtained through meta-training is loaded and kept fixed without updating. The parameters of the cross-modal fusion model are used to perform fusion processing on the data in the small sample to generate scale prediction fusion features. Load the parameters of the procurement scale prediction model obtained by training, use the parameters of the procurement scale prediction model to perform prediction processing on the scale prediction fusion features, compare the prediction results with the actual procurement scale in the small sample, and calculate the small sample prediction loss. Based on the small sample prediction loss, gradient descent is performed to update the parameters of the procurement scale prediction model, resulting in updated procurement scale prediction model parameters. These updated parameters will replace the original model parameters for subsequent real-time predictions.

[0081] The actions of freezing the backbone and fine-tuning the prediction head in this step are completely consistent with the simulated actions of the "inner loop stage" in meta-training. This indicates that the meta-training stage has successfully trained the backbone and mastered the ability to extract easily adaptable features. This allows the meta-learning results to be obtained in the most efficient way (fine-tuning the prediction head only) in the practical application of step 5.

[0082] This step enables adaptive dynamic updates of the model during the real-time prediction phase, ensuring that the model maintains prediction accuracy and stability even under environmental changes, data distribution shifts, or changes in business structure, thereby improving long-term effectiveness and robustness.

[0083] Example 2: A power bidding and procurement scale prediction system, which implements a power bidding and procurement scale prediction method, includes: The data acquisition and preparation module is used to acquire historical power bidding and procurement data, including unstructured text data representing technical requirements and structured data representing the business operation environment. A cross-modal model building module is used to establish a cross-modal fusion model, which performs fusion processing on the unstructured text data and the structured data to generate scale prediction fusion features; and to establish a procurement scale prediction model, which performs prediction processing on the scale prediction fusion features to obtain the power bidding procurement scale. The meta-training and parameter generation module is used to divide the historical power bidding and procurement data into multiple meta-learning task sets according to the preset power bidding and procurement scenario; using a meta-learning framework, based on the multiple meta-learning task sets, meta-training is performed on the cross-modal fusion model parameters and procurement scale prediction model parameters to obtain updated cross-modal fusion model parameters and procurement scale prediction model parameters. The real-time forecasting module is used to acquire the power bidding and procurement data to be predicted, and based on the loaded and updated cross-modal fusion model parameters and procurement scale forecasting model parameters, it performs fusion processing and forecasting processing to obtain the power bidding and procurement scale. The online monitoring and adaptive update module is used to monitor whether the completed power bidding and procurement data has triggered preset conditions; if the preset conditions are triggered, the completed power bidding and procurement data and the corresponding actual procurement scale are used as a small sample to update the parameters of the procurement scale prediction model.

[0084] Example 3: A terminal, comprising a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method.

[0085] Example 4: A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method.

[0086] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0087] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0088] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0089] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for predicting the scale of power solicitation procurement, characterized by, The method comprises the following steps: Step 1, obtaining historical power bidding procurement data; Step 2, establishing a cross-modal fusion model and a procurement scale prediction model, respectively used for generating scale prediction fusion features and outputting power bidding procurement scale; Step 3, dividing the historical power bidding procurement data into multiple meta-learning task sets, and performing meta-training on the cross-modal fusion model parameters and the procurement scale prediction model parameters to obtain updated cross-modal fusion model parameters and procurement scale prediction model parameters; Step 4, obtaining the power bidding procurement data to be predicted, and obtaining the power bidding procurement scale based on the loaded updated cross-modal fusion model parameters and procurement scale prediction model parameters; Step 5, when a completed power bidding procurement data triggers a preset condition, the completed power bidding procurement data and the corresponding real procurement scale are used as a small sample to update the procurement scale prediction model parameters.

2. The method of claim 1, wherein, The historical power bidding procurement data comprises: unstructured text data representing technical requirements, including power project technical requirement documents; structured data representing business operation environment, including category type data and numerical type data; wherein the category type data includes engineering project attributes, power grid load prediction level, and seasonal feature labels; the numerical type data includes bulk commodity price index; and historical real procurement scale data corresponding to the unstructured text data and the structured data.

3. The method of claim 2, wherein, In step 2, the cross-modal fusion model comprises: a text encoder for performing word segmentation processing on the unstructured text data to generate a token sequence; inputting the token sequence into a pre-trained language model to generate a technical requirement embedding vector; a structured data encoder for performing entity embedding processing on the category type data in the structured data to generate a category feature vector; concatenating the category feature vector with the numerical type data in the structured data, and inputting them into a multi-layer perceptron for nonlinear transformation to generate an operation environment embedding vector; a cross-modal fusioner for receiving and fusing the technical requirement embedding vector and the operation environment embedding vector to obtain the scale prediction fusion features.

4. The method of claim 3, wherein, The structure of the cross-modal fusioner comprises: a concatenation and embedding layer for receiving the technical requirement embedding vector and the operation environment embedding vector; concatenating the technical requirement embedding vector and the operation environment embedding vector in the sequence dimension to form a unified vector sequence, and adding position encoding and modal type encoding to the unified vector sequence; N-layer Transformer encoder layers for receiving the output of the concatenation and embedding layer, and each layer comprising a multi-head self-attention sublayer and a feedforward network sublayer; wherein the multi-head self-attention sublayer is used for performing self-attention calculation on the unified vector sequence; the feedforward network sublayer is used for performing nonlinear transformation on the output of the multi-head self-attention sublayer; a feature pooling layer configured to receive a final unified vector sequence output by a last layer of the N-layer transformer encoder layer, and perform a pooling operation on the final unified vector sequence to generate and output the scale prediction fusion feature.

5. The method of claim 3, wherein, In step 2, the structure of the scale prediction model includes: an input layer configured to receive the scale prediction fusion feature output by the cross-modal fusion model; a hidden layer having a nonlinear activation function configured to perform a nonlinear transformation on the scale prediction fusion feature; and an output layer having a neuron with linear activation to obtain the power tendering scale.

6. The method of claim 1, wherein, In step 3, the cross-modal fusion model parameters and the scale prediction model parameters are meta-trained based on a plurality of meta-learning task sets, including: dividing each meta-learning task in the plurality of meta-learning task sets into a support set and a query set; sampling M meta-learning tasks from the plurality of meta-learning task sets for iterative training, the iterative training including an inner loop phase and an outer loop phase; in the inner loop phase, for each meta-learning task in the M meta-learning tasks, loading the current scale prediction model parameters, calculating a loss using the corresponding support set, and performing gradient descent update on the scale prediction model parameters to obtain task-specific parameters; in the outer loop phase, using the task-specific parameters and the current cross-modal fusion model parameters to calculate a prediction loss on the corresponding query set, and accumulating the prediction losses of the M meta-learning tasks to obtain a meta loss; based on the meta loss, simultaneously performing gradient descent update on the cross-modal fusion model parameters and the scale prediction model parameters to obtain updated cross-modal fusion model parameters and scale prediction model parameters.

7. The method of claim 1, wherein, In step 5, the completed power tendering data and the corresponding true scale are used as a small sample to update the scale prediction model parameters, including: loading the cross-modal fusion model parameters and the scale prediction model parameters obtained by meta-training; using the cross-modal fusion model parameters to perform fusion processing on the data in the small sample to generate scale prediction fusion features; using the scale prediction model parameters to perform prediction processing on the scale prediction fusion features, and comparing the prediction results with the true scale in the small sample to calculate a small sample prediction loss; based on the small sample prediction loss, performing gradient descent update on the scale prediction model parameters to obtain updated scale prediction model parameters.

8. A power tender procurement scale prediction system that operates a power tender procurement scale prediction method as claimed in any one of claims 1 to 7, characterized by, including: a data acquisition and preparation module configured to acquire historical power tendering data, including unstructured text data representing technical requirements and structured data representing business operation environment; a cross-modal model construction module configured to establish a cross-modal fusion model configured to perform fusion processing on the unstructured text data and the structured data to generate scale prediction fusion features; and a scale prediction model configured to perform prediction processing on the scale prediction fusion features to obtain the power tendering scale. The meta-training and parameter generation module is configured to divide the historical power bidding and purchasing data into a plurality of meta-learning task sets according to a preset power bidding and purchasing scenario; The meta-learning framework is adopted to perform meta-training on the cross-modal fusion model parameters and the purchasing scale prediction model parameters based on the plurality of meta-learning task sets, so as to obtain updated cross-modal fusion model parameters and purchasing scale prediction model parameters; The real-time prediction module is configured to acquire power bidding and purchasing data to be predicted, and perform fusion processing and prediction processing based on the updated cross-modal fusion model parameters and the purchasing scale prediction model parameters, so as to obtain a power bidding and purchasing scale. The online monitoring and adaptive updating module is configured to monitor whether the completed power bidding and purchasing data triggers a preset condition. In a case where the preset condition is triggered, the completed power bidding and purchasing data and a corresponding real purchasing scale are used as a small sample to update the purchasing scale prediction model parameters.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is configured to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method according to any one of claims 1-7.

Images