Investment optimization calculation method and device based on multi-modal artificial intelligence model, equipment and medium

Abstract

The application relates to an investment optimization calculation method and device based on a multi-modal artificial intelligence model, equipment and a medium, which comprises the following steps: performing unified format processing on the integrated first multi-source data to form a multi-modal feature vector that can be directly input into a pre-trained model; obtaining prediction values of power grid project construction duration, settlement duration and final settlement duration and a monthly investment proportion sequence based on the multi-modal feature vector; adjusting the investment proportion of the remaining months in real time; based on the updated monthly investment proportion, reallocating annual investment quotas among multiple power grid projects according to a fund utilization rate threshold; and generating annual investment quotas, milestone plans and fund expenditure curves of a single power grid project according to the adjusted annual investment plan. The application enhances the prediction capability of the model and the reliability of decision support, improves the adaptability and flexibility of investment decisions, and improves the adaptability and flexibility of investment plans.

Description

Technical Field

[0001] This application relates to the technical field of power grid investment calculation, and in particular to investment optimization calculation methods, devices, equipment and media based on multimodal artificial intelligence models. Background Technology

[0002] Investment decisions for power grid projects directly impact the stability and economic efficiency of power supply. Traditional investment calculation methods primarily rely on historical data and human experience, employing static models such as fuzzy analysis, analytic hierarchy process (AHP), and grey relational analysis, which have significant technical limitations. First, the problem of data silos is prominent; data from planning, infrastructure, and ERP systems are not effectively integrated, leading to insufficient decision-making basis. Second, the models lack dynamic adjustment capabilities, making it difficult to cope with delays or deviations in investment progress caused by changes in the construction environment (such as weather and geographical location). Finally, the processing capabilities for unstructured data (such as images and text) are insufficient, failing to fully utilize multi-source information to improve prediction accuracy.

[0003] The existing power grid investment calculation technology has the following main shortcomings: 1. Insufficient data integration: Data from different systems is not interconnected, creating information silos and hindering collaborative analysis of multi-source data. 2. Static models: Relying on historical data and fixed parameters, they lack the ability to dynamically respond to construction environments (such as weather and geographical location). 3. Weak unstructured data processing capabilities: Multimodal data such as images and text are not effectively integrated into the investment decision-making process. 4. Lack of automation mechanisms: Model building relies on manual selection and optimization, making it difficult to meet the investment optimization needs of large-scale project groups. 5. Lack of effective closed loop: The calculation results and implementation do not form an effective closed loop, resulting in a disconnect.

[0004] Therefore, there is an urgent need for a machine learning-driven investment optimization calculation scheme for power grid projects. By integrating structured and unstructured data (such as documents, images, and geographic information) of power grid projects, this scheme can utilize automated machine learning technology to achieve high-precision prediction and dynamic adjustment of investment plans, providing intelligent investment decision support for power grid companies. Summary of the Invention

[0005] In order to integrate multi-source data and realize real-time economic optimization of power grid project investment calculation, this application provides an investment optimization calculation method, device, equipment and medium based on a multimodal artificial intelligence model.

[0006] Firstly, this application provides an investment optimization calculation method based on a multimodal artificial intelligence model, employing the following technical solution: An investment optimization calculation method based on a multimodal artificial intelligence model includes: Acquire first multi-source data for the power grid project, which includes structured data, unstructured data, and environmental parameters. The structured data includes at least the planned start date, planned commissioning date, planned total investment, total investment approved in the feasibility study, approval date, line length, and substation capacity. The unstructured data includes at least the feasibility study report text, construction service contract text, material procurement contract text, construction progress images, GIS topology, and DEM elevation data. The environmental parameters include at least historical and predicted meteorological time series and geographic location topology data of the construction area. The integrated first multi-source data is uniformly formatted to form a multimodal feature vector that can be directly input into the pre-trained model; The multimodal feature vectors are input into the trained multimodal construction period calculation model and monthly investment amount calculation model respectively to obtain the predicted values ​​of the construction period, settlement period, and final account period of the power grid project, as well as the monthly investment ratio sequence. The system collects real-time progress images and the latest meteorological data from the construction site, calculates the deviation between the actual progress and the predicted progress, and incrementally updates the construction period calculation model and the monthly investment amount calculation model through a continuous learning interface, adjusting the investment ratio for the remaining months in real time. Based on the updated monthly investment ratio, the annual investment amount is redistributed among multiple power grid projects according to the capital utilization rate threshold to achieve cross-power grid project capital balance. Based on the adjusted annual investment plan, the annual investment amount, milestone plan and capital expenditure curve of a single power grid project are generated, which serve as the final output of the power grid company's intelligent investment decision-making.

[0007] By adopting the above technical solutions, multi-source data can be integrated to achieve real-time economic optimization calculations for power grid project investment. This enhances the model's predictive capabilities and the reliability of decision support, improves the adaptability and flexibility of investment decisions, and enhances the adaptability and flexibility of investment plans.

[0008] Optionally, the training methods for the multimodal construction period calculation model and the monthly investment amount calculation model include: The merged structured data, unstructured data, and time series data are used to construct a unified dataset; Using AutoGluon's MultiModalPredictor, Transformer, CNN, and time series models are automatically searched and integrated on the unified dataset to perform multi-label regression training on construction period, settlement period, and final account period, resulting in a multimodal period calculation model. Based on the multimodal construction period calculation model, using AutoGluon's TimeSeriesPredictor, SeasonalNaive, ETS, DeepAR, and PatchTST time series models are automatically combined. With the construction period prediction value, static characteristics of the power grid project, historical investment, construction progress, meteorological information, and DEM elevation data as inputs, a monthly investment ratio sequence is generated to obtain a monthly investment amount calculation model. Based on cross-validation and automatic hyperparameter tuning, the model structure and weights of the multimodal construction period calculation model and the monthly investment amount calculation model are determined, and the model accuracy is verified on the test set. Based on the continuous learning framework, construction deviation data is received in real time, and the multimodal construction period calculation model and the monthly investment amount calculation model are incrementally updated.

[0009] By adopting the above technical solutions, the annual investment plans for power grid projects are predicted, and the prediction performance is improved through automated hyperparameter tuning. Combining the AutoGluon continuous learning framework, a real-time feedback interface is designed to dynamically update the model with new data such as construction progress deviations, enabling adaptive adjustments to the investment plan.

[0010] Optionally, the uniform formatting process includes: The structured data is subjected to numerical processing, null padding, multicollinearity removal, and logarithmic standardization. Based on a pre-defined large language model, the necessity of construction, construction scale, construction requirements, and material supply requirements are extracted from the unstructured data and tagged accordingly. Based on a preset visual model, key equipment and construction stages are identified from the construction progress images and output with semantic labels. For the GIS topology and DEM data, extract the terrain type, path length and terrain complexity of the construction path and generate terrain feature vectors; For meteorological time series, extract the number of extreme weather days and generate the number of days of work stoppage due to weather impact; Based on a unified CSV or tensor format, the processed first multi-source data is aligned and serialized to form the multimodal feature vector that can be directly input into the pre-trained model.

[0011] By adopting the above technical solutions, data barriers are broken down, enabling efficient fusion of multi-source heterogeneous data. This improves the adaptability and performance ceiling of pre-trained models.

[0012] Optionally, the method further includes: Secondary multi-source data for the corresponding area of ​​the project is acquired, and spatiotemporal drift detection processing is performed on the secondary multi-source data to obtain the spatiotemporal drift detection results; the secondary multi-source data includes mobile phone signaling grid data, meteorological monitoring data, power grid operation data, and disaster early warning data; Based on the spatiotemporal drift detection results, a load impact factor matrix is ​​generated; the load impact factor matrix includes population characteristics, meteorological deviation, and disaster attenuation coefficient. Based on the load influence factor matrix and historical load sequence, a load prediction hyperbola is generated through a pre-trained conditional quantile autoregressive model; the historical load sequence is the active power time series data recorded by power grid equipment at various historical moments, and the load prediction hyperbola includes a median curve and an upper limit curve. Based on the comparison results of the load forecast hyperbola and the dynamic threshold, it is determined whether to trigger a power grid project investment decision; the comparison results include the project risk level and the project power shortage parameters. If so, then based on the preset economic optimization model, the predicted net present value and internal rate of return corresponding to the power grid project investment are calculated; the predicted net present value is a parameter characterizing the absolute return of the power grid project investment, and the internal rate of return is a parameter characterizing the relative return of the power grid project investment. If the predicted net present value is non-negative and the internal rate of return is not lower than the adjusted discount rate, then an expansion investment instruction is output; the expansion investment instruction includes the target expansion plan and the required equipment list.

[0013] By adopting the above technical solutions, four types of data—mobile phone signaling grid data, meteorological monitoring data, power grid operation data, and disaster early warning data—are automatically aligned within a four-dimensional spatiotemporal grid, reducing drift positioning errors, minimizing abnormal peak values, and achieving the fusion of multi-source heterogeneous data. The conditional quantile autoregressive model outputs median and upper limit curves, providing conservative boundaries, reducing the possibility of underestimating extreme loads, enabling the generation of real-time load forecast hyperbolas, and, based on an economic optimization model, realizing real-time economic optimization calculations for power grid project investment.

[0014] Optionally, before determining whether to trigger a power grid project investment decision based on the comparison results of the load forecast hyperbola and the dynamic threshold, the method further includes: Based on the historical load sequence and the historical load forecast sequence, the historical monthly forecast error is calculated, and based on the historical monthly forecast, the basic threshold for monthly updates is obtained; the historical load forecast sequence is the forecast result sequence of the conditional quantile autoregressive model for the load at historical time, and the historical monthly forecast error is the average forecast deviation of the conditional quantile autoregressive model within the monthly cycle. Based on real-time prediction error, meteorological temperature difference, homecoming index, and dummy variables of emergencies, a real-time factor vector is calculated. This real-time factor vector includes an error factor, a holiday factor, and an event factor. The real-time prediction error is the instantaneous deviation at the current moment; the meteorological temperature difference is the difference between the current temperature and the historical average temperature for the same period; the homecoming index is a parameter characterizing population migration within the project's corresponding area; and emergencies include disasters or regional power outages occurring within the project's corresponding area. Based on a trained LSTM network model, a weight vector corresponding to the real-time factor vector is obtained. This weight vector includes the weights corresponding to the error factor, the holiday factor, and the event factor. The dynamic threshold is obtained based on the base threshold, the real-time factor vector, and the weight vector.

[0015] By adopting the above technical solutions, the second multi-source data is aligned through a four-dimensional spatiotemporal grid, the drift marker is verified by conflict resolution rules, and the drift score is output by a scenario weighting mechanism. Furthermore, the dynamic scenario weighting makes the load forecast more in line with actual needs.

[0016] Optionally, after the output expansion investment instruction, the following may also be included: If the current power grid project is successfully put into operation, the operation data after the power grid project is collected in real time; the operation data includes the privacy-preserving actual load rate, actual investment amount, and actual net present value; the privacy is achieved by applying differential privacy noise to the operation data; Based on the operational data, the expansion success rate index is calculated, and based on the expansion success rate index, the project level corresponding to the current power grid project is obtained. Based on the project level, the initial benchmark discount rate for subsequent power distribution capacity expansion investment projects will be dynamically adjusted.

[0017] By adopting the above technical solution, real-time collection of privacy-protected operational data and calculation of expansion success rate index are used to quantify the operation results into project level, realizing a complete closed loop of pre-investment prediction, post-investment verification, and model retraining. The initial benchmark discount rate of subsequent projects is automatically adjusted up or down according to the project level, so that the capital cost of expansion investment is aligned with the real risk in real time.

[0018] Optionally, before the output expansion investment instruction, the following may also be included: Based on the long-term planning data of the corresponding area of ​​the project, the current capacity expansion plan is validated under a scenario; the long-term planning data includes spatial planning data, industrial relocation plan data, and distributed energy access planning data. If the scenario verification result indicates that the capacity expansion requirement corresponding to the current capacity expansion plan is redundant within the long-term period, then a capacity tiered decision-making mechanism is initiated. The capacity tiered decision-making mechanism includes: using the difference between the medium-term load forecast curve and the long-term load forecast curve as the redundant capacity, and using the redundant capacity as a constraint to re-optimize the current capacity expansion plan and generate a tiered capacity expansion path. The tiered capacity expansion path includes the initial capacity expansion, medium-term scalable interfaces, and long-term reserved capacity. The medium-term load forecast curve and the long-term load forecast curve are obtained by the conditional quantile autoregressive model using the spatial planning data as a new exogenous variable and re-rolling the forecast, with the forecast step size corresponding to the medium-term planning period and the long-term planning period, respectively. Based on the tiered capacity expansion path, update the static investment cost, expected unloaded cost, and target benchmark discount rate, and recalculate the predicted net present value and the internal rate of return. If the updated predicted net present value is still non-negative and the internal rate of return is not lower than the adjusted discount rate, then the target capacity expansion plan is generated based on the current capacity expansion plan and the tiered capacity expansion path.

[0019] By adopting the above technical solution, the sunk cost of redundant capacity is reduced and the risks of capital occupation and asset idleness are reduced through a three-stage capacity ladder of initial, medium and long term.

[0020] Secondly, this application provides an investment optimization calculation device based on a multimodal artificial intelligence model, which adopts the following technical solution: An investment optimization calculation device based on a multimodal artificial intelligence model includes: The data acquisition module is used to acquire the first multi-source data of the power grid project. The first multi-source data includes structured data, unstructured data, and environmental parameters. The structured data includes at least the planned start date, planned commissioning date, planned total investment, total investment approved in the feasibility study, approval date, line length, and substation capacity. The unstructured data includes at least the feasibility study report text, construction service contract text, material procurement contract text, construction progress images, GIS topology, and DEM elevation data. The environmental parameters include at least the historical and predicted meteorological time series and geographical location topology data of the construction area. The format processing module is used to uniformly format the integrated first multi-source data to form a multimodal feature vector that can be directly input into the pre-trained model. The model prediction module is used to input the multimodal feature vectors into the trained multimodal construction period calculation model and the monthly investment amount calculation model respectively, so as to obtain the predicted values ​​of the construction period, settlement period, and final account period of the power grid project, as well as the monthly investment ratio sequence. The deviation calculation module is used to collect progress images and the latest meteorological data at the construction site in real time, calculate the deviation between the actual progress and the predicted progress, and incrementally update the construction period calculation model and the monthly investment amount calculation model through the continuous learning interface, and adjust the investment ratio of the remaining months in real time. The reallocation module is used to reallocate annual investment amounts among multiple power grid projects based on the updated monthly investment ratio and according to the capital utilization threshold, so as to achieve cross-power grid project capital balance. The decision output module is used to generate the annual investment amount, milestone plan and capital expenditure curve of a single power grid project based on the adjusted annual investment plan, and serves as the final output of the power grid company's intelligent investment decision-making.

[0021] Thirdly, this application provides an electronic device that adopts the following technical solution: An electronic device includes a processor and a memory, wherein the processor is coupled to the memory; The processor is configured to execute a computer program stored in the memory, causing the electronic device to perform the method as described in any of the first aspects.

[0022] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium includes a computer program or instructions that, when executed on a computer, cause the computer to perform the method as described in any of the first aspects. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an investment optimization calculation method based on a multimodal artificial intelligence model, according to one embodiment of this application.

[0024] Figure 2 This is a flowchart illustrating the multimodal data fusion process of one embodiment of this application.

[0025] Figure 3 This is a flowchart illustrating the Autogluon hybrid model construction process of one embodiment of this application.

[0026] Figure 4 This is a flowchart illustrating the auxiliary investment decision-making and dynamic optimization process of one embodiment of this application.

[0027] Figure 5 This is a flowchart illustrating a continuous optimization model of one embodiment of this application.

[0028] Figure 6 This is a structural block diagram of an investment optimization calculation device based on a multimodal artificial intelligence model, according to one embodiment of this application.

[0029] Figure 7 This is a structural block diagram of an electronic device according to one embodiment of this application. Detailed Implementation

[0030] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0031] The present application will be further described in detail below with reference to the accompanying drawings.

[0032] This application provides an investment optimization calculation method based on a multimodal artificial intelligence model. This method can be executed by a device, which can be a server or a terminal device. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal device can be a smartphone, tablet, desktop computer, etc., but is not limited to these.

[0033] Example 1 like Figure 1 As shown, an investment optimization calculation method based on a multimodal artificial intelligence model, with electronic devices as the execution subject, is described in its main process flow as follows (steps S101 to S106): Step S101: Obtain the first multi-source data for the power grid project. The first multi-source data includes structured data, unstructured data, and environmental parameters. The structured data includes at least the planned start date, planned commissioning date, planned total investment, total investment approved in the feasibility study, approval date, line length, and substation capacity. The unstructured data includes at least the feasibility study report text, construction service contract text, material procurement contract text, construction progress images, GIS topology, and DEM elevation data. The environmental parameters include at least the historical and predicted meteorological time series and geographical location topology data of the construction area. Step S102: Perform unified formatting on the integrated first multi-source data to form a multimodal feature vector that can be directly input into the pre-trained model; Step S103: Input the multimodal feature vectors into the trained multimodal construction period calculation model and monthly investment amount calculation model respectively to obtain the predicted values ​​of the construction period, settlement period, and final account period of the power grid project, as well as the monthly investment ratio sequence; Step S104: Collect progress images and the latest meteorological data at the construction site in real time, calculate the deviation between the actual progress and the predicted progress, and incrementally update the construction period calculation model and the monthly investment amount calculation model through the continuous learning interface, and adjust the investment ratio of the remaining months in real time; Step S105: Based on the updated monthly investment ratio, reallocate the annual investment amount among multiple power grid projects according to the capital utilization rate threshold to achieve cross-power grid project capital balance; Step S106: Based on the adjusted annual investment plan, generate the annual investment amount, milestone plan and capital expenditure curve for a single power grid project, and use it as the final output of the power grid company's intelligent investment decision-making.

[0034] In this embodiment, the training method for the multimodal construction period calculation model and the monthly investment amount calculation model includes: constructing a unified dataset from the fused structured data, unstructured data, and time series data; using AutoGluon's MultiModalPredictor, automatically searching and integrating Transformer, CNN, and time series models on the unified dataset, and performing multi-label regression training on the construction period, settlement period, and final account period to obtain the multimodal construction period calculation model; based on the multimodal construction period calculation model, using AutoGluon's TimeSeriesPredictor, The system automatically combines SeasonalNaive, ETS, DeepAR, and PatchTST time series models, using project duration forecasts, static characteristics of power grid projects, historical investments, construction progress, meteorological information, and DEM elevation data as inputs to train and generate monthly investment ratio sequences, thus obtaining a monthly investment amount calculation model. Based on cross-validation and automatic hyperparameter tuning, the model structure and weights of the multimodal project duration calculation model and the monthly investment amount calculation model are determined, and the model accuracy is verified on a test set. Based on a continuous learning framework, construction deviation data is received in real time, and the multimodal project duration calculation model and the monthly investment amount calculation model are incrementally updated.

[0035] In this embodiment, the unified formatting process includes: performing numerical conversion, null value filling, multicollinearity removal, and logarithmic standardization on the structured data; extracting and labeling construction necessity, construction scale, construction requirements, and material supply requirements from the unstructured data based on a preset large language model; identifying key equipment and construction stages from the construction progress image based on a preset visual model and outputting them with semantic labels; extracting terrain category, path length, and terrain complexity of the construction path from the GIS topology and DEM data and generating terrain feature vectors; extracting the number of extreme weather days from the meteorological time series and generating the number of days of work stoppage due to weather impact; and aligning and serializing the processed first multi-source data based on a unified CSV or tensor format to form the multimodal feature vector that can be directly input into the pre-trained model.

[0036] Specifically: Multimodal data fusion methods specifically include: Data standardization: Classified text, meteorological data, and geographic latitude and longitude are numerically processed using methods such as one-hot encoding and text embedding; logarithmic calculation is used to normalize and standardize numerical data to eliminate the impact of dimensional differences on model feature weights; and open-source models such as PaddleClas are used to label infrastructure construction images and identify construction progress for numerical processing.

[0037] Cross-modal joint modeling: Using AutoGluon's MultiModalPredictor module, structured data, unstructured data, and time series data (meteorological records) are integrated into a unified dataset to achieve automated extraction and fusion of multimodal features.

[0038] Data input format: Table data and image paths are linked through a CSV file, and meteorological time series data are embedded in the table in a multi-column format to ensure data compatibility.

[0039] The Autogluon hybrid model construction technology solution specifically includes: Model selection and training: First, MultiModalPredictor is used to automatically select models suitable for power grid investment scenarios (such as integrating Transformer, CNN and time series models) to build a multimodal project duration calculation model; second, based on the results of the project duration prediction model, a time series prediction model is built to predict the possible investment amount of each power grid project in each month, thereby predicting the annual investment plan of the power grid project; and prediction performance is improved by using automated hyperparameter tuning.

[0040] Dynamic environmental parameter integration: After the meteorological time series is preprocessed by the TimeSeriesPredictor module, the number of extreme weather days that affect construction (such as the number of days when excavation cannot be carried out due to frozen soil) is extracted and input into the model together with the structured data; the geographic location information is analyzed by combining GIS topology with DEM model data to analyze the different types of terrain and corresponding path lengths through which construction is carried out, and is transformed into feature vectors (such as path length and terrain complexity).

[0041] Iterative optimization mechanism: Combining AutoGluon's continuous learning framework, a real-time feedback interface is designed to dynamically update the model with new data such as construction progress deviations, thereby enabling adaptive adjustments to the investment plan.

[0042] like Figure 2 As shown, the power grid multimodal data fusion process specifically includes: The processing of structured data includes: using visualization data analysis tools such as seaborn to analyze and process structured data related to the construction period, such as planned start date, planned commissioning date, planned total investment, feasibility study approval total investment, and approval date. Numerical processing: For some categorical attributes, numerical processing is performed using dictionary encoding or one-hot encoding; Null value handling: Different strategies such as mean, median, mode, and zero value are used to fill null values ​​for different attributes; Multicollinearity handling: Multicollinearity is measured using the variance inflation factor (existing technology, not described in detail here) to remove some multicollinearity attributes; Single-value attribute processing: View the distribution histogram of each attribute and remove single-value features.

[0043] Correlation analysis: Plot the regression fit scatter plot and heatmap of each attribute relative to the predicted value, analyze the correlation of each feature attribute with the prediction result, and remove attributes with low correlation.

[0044] Large numerical attribute processing: For attributes with huge individual numerical differences, such as total investment of power grid projects and line length, standardization or logarithmic methods are used for processing. After consideration, this application can use logarithmic processing.

[0045] After the above processing, the retained attribute features are used for subsequent construction of the multimodal project duration calculation model, mainly including the following information: Voltage level, affiliated unit, construction nature, asset attributes, whether feasibility study review has been completed, whether feasibility study approval has been completed, whether verification has been completed, total investment (take ln, reduce the larger value), number of lines, total line length (take ln), overhead line length (take ln), cable length (take ln), number of substations, substation capacity (take ln), whether preliminary design approval has been completed, functional category, power supply area, number of bays, number of substations, number of low-voltage overhead lines, length of low-voltage overhead lines (take ln), number of low-voltage cables, length of low-voltage cables (take ln), number of low-voltage supporting lines, length of low-voltage supporting lines (take ln), number of optical cables, length of optical cables (take ln), planned construction period (number of months from planned commissioning date to planned start date), number of months from preliminary design approval to planned commissioning, number of months from feasibility study review to planned commissioning, number of months from feasibility study approval to planned commissioning, number of months from verification to planned commissioning, month of commencement, etc.

[0046] Unstructured text data such as feasibility study reports, construction service contracts, and material procurement contracts are input into an LLM (Local Management Model) to extract key information such as construction necessity, construction scale, construction requirements, and material supply requirements. Based on the extracted information, the LLM is used to label the projects according to power grid project classification and urgency. After handling null values, the extracted key information and labels are used together in a multimodal project duration estimation model. This mainly includes: Necessity of construction, contract amount (ln), construction period (if any) in the construction contract, payment terms in the construction contract, contract amount (ln) for material procurement, delivery time (if any) for material procurement, payment terms in the material procurement contract, etc.

[0047] Using PaddleClas or a multimodal large model, key content (such as transformers, towers, etc.) is identified in construction images, and the image content is described semantically. After null value processing, the extracted semantic descriptions are retained for subsequent use in the multimodal project duration calculation model. These descriptions mainly include: whether equipment has arrived on site, civil engineering progress descriptions, and power line erection progress descriptions.

[0048] Extract paths from the GIS topology of the power grid project to obtain the construction paths. Combined with DEM information from various locations, extract the terrain and detailed attributes (such as hill height) involved in the power grid project construction. After null value processing, the extracted terrain data is retained for subsequent use in the multimodal project duration calculation model. This mainly includes: whether there is work in mountainous areas, the slope of mountainous areas (ln), whether there is work across rivers, and the length of the river crossing (ln).

[0049] Collect expert experience data on power grid construction projects in various provinces to obtain extreme weather conditions affecting construction (such as temperature levels that would trigger work stoppages). Collect historical and forecast meteorological data from various regions to predict the number of days of work stoppages due to weather conditions (relevant forecast data can be dynamically collected during construction and used as dynamic covariates in the calculations). Use the predicted data for a multimodal construction period calculation model. This mainly includes: the number of extreme high-temperature days, extreme low-temperature days, extreme high-wind days, and typhoon days.

[0050] like Figure 3 As shown, the construction process of the multimodal project duration estimation model and the monthly investment plan estimation model specifically includes: Data loader: Classifies and loads text data, tabular (numerical) data, and time-series data. Based on the actual start date, actual commissioning date, actual settlement date, and actual final settlement date of historical power grid projects, it calculates three target values ​​for construction period, settlement period, and final settlement period as target labels for the calculation model.

[0051] Sample splitting: The sample data is divided into training and test sets in a 1:1 ratio. Multimodal project duration estimation model construction: A multimodal project duration estimation model was built based on Autogluon's MultiModalPredictor, and iteratively trained on the training set with three target labels in multiple rounds. Autogluon automatically selects and merges the text model and regression model, calculates the average loss, and automatically tunes the hyperparameters and weight parameters of each model. The model accuracy was verified on the test set until it met the experimental accuracy requirements.

[0052] Monthly Investment Calculation Model Construction: Combining the calculated construction period, static characteristics of the power grid project, historical investment already transmitted, historical construction progress (time-varying covariate), and meteorological information (time-varying covariate), a monthly investment calculation model is constructed using Autogluon's TimeSeriesPredictor, which automatically combines multiple time series models (such as SeasonalNaiveModel, ETSModel, DeepARModel, PatchTSTModel, etc.). The model is trained on the training set and evaluated on the test set until the experimental accuracy requirements are met. Special processing notes: All power grid projects are grouped according to their affiliated unit, voltage level, project type, start month, and construction period. This assumes that power grid projects with the same unit, voltage, type, start month, and construction period exhibit consistent theoretical investment patterns. This group is then used as a small dataset for training, producing a sub-hybrid model. During prediction, the corresponding sub-model is selected based on the unit, voltage, type, start month, and construction period of the power grid project to be predicted.

[0053] Because the actual investment start years for power grid projects vary, to ensure consistency of input features, the actual year is not considered during model training. Only the month in which the power grid project starts is considered (the start month affects the number of months available for construction that year, and the weather during the start season also affects the duration of construction). The monthly investment ratio (relative to the total investment) of the sample power grid projects from the start month is used as the time series feature. To ensure that the model sufficiently grasps the theoretical monthly investment data for the grouped power grid projects, two repeated time series periods are added for each power grid project.

[0054] Assuming a power grid project has a construction period of 3 months, the following table shows an example of a time column with two repeating cycles: Table 1 Following the example above, when predicting a power grid project with a 3-month construction period, it is only necessary to predict when the investment ratio first reaches 1.0 (i.e., 100%) or when the investment ratio decreases significantly compared to the previous month. This allows for repeated learning of the monthly investment sequence pattern during training without affecting subsequent practical calculations.

[0055] Application of the monthly investment calculation model: Based on the power grid project's affiliated unit, voltage level, project type, start month, and construction period, an appropriate sub-model is selected to predict the investment ratio for each month after commencement. Then, each monthly ratio is multiplied by the total investment to obtain the monthly investment amount. For newly started power grid projects, the data for the nth month (denoted as n) is extracted from the planned start date (or actual start date, if any) to the end of December of the current year. The same operation applies to ongoing or completed power grid projects (starting from January of the current year and ending at the end of December, or up to the planned final settlement date).

[0056] Output of calculation results for a single power grid project: construction period, settlement period, final settlement period, and monthly investment.

[0057] like Figure 4 As shown, the process of supporting investment decision-making and dynamic optimization specifically includes: Construction period calculation: Based on the planned commissioning time and the need for safe supply, select power grid projects with higher urgency and use a multimodal construction period calculation model to calculate the construction period of the power grid projects.

[0058] Define milestone plans: Determine the start or production date, and derive milestone plan information such as planned start date, planned production date, planned settlement date, and planned final account date by tracing forward or backward.

[0059] Calculate monthly investment amount: Combining data such as planned milestone information, use the monthly investment amount calculation model to calculate the monthly investment amount of the power grid project.

[0060] Calculate the annual investment plan: Based on the planned start date of each power grid project, calculate the number of construction months for each power grid project in the current year, and combine this with the monthly investment amount to calculate the annual investment plan for each power grid project.

[0061] Company investment decisions: Taking into account boundary factors such as the investment capacity of each unit, the annual investment plans of each power grid project are summarized, and the investment plan of the branch company is derived by balancing them.

[0062] Plan Execution and Dynamic Optimization: Monitor dynamic environmental characteristics such as power grid project maturity, meteorological information, and construction images; dynamically calculate the construction period and annual investment plan for each power grid project; if the calculation results exceed a preset threshold, automatically trigger adjustments to the annual plan. Summarize the adjustments for all power grid projects to derive the branch company's adjustment plan.

[0063] like Figure 5 As shown, the process of continuously optimizing the model specifically includes: New Sample Collection: Collect data from newly commissioned power grid projects as new samples to supplement the training / test set; Model Training: Conduct model training and evaluation based on the new dataset, and output the optimized model; Optimization Decision: Conduct calculations and applications based on the new model.

[0064] In this embodiment, by fusing structured and unstructured data, efficient processing and integration of multimodal data are achieved, enhancing the accuracy and reliability of the prediction model. Environmental parameters such as geographical location and construction weather are introduced to construct a dynamic feedback mechanism, adjusting investment plans in real time and improving the adaptability of investment decisions. Autogluon is used to automate the construction and hyperparameter tuning of the hybrid model, reducing manual intervention and improving model construction efficiency and prediction accuracy. Through continuous model optimization, dynamic optimization is achieved throughout the entire process from data collection and model training to investment decision-making, forming an effective closed-loop feedback mechanism and improving the adaptability and flexibility of the investment plan.

[0065] Specifically: The adoption of multi-source data fusion technology enhances the comprehensiveness and accuracy of decision-making data, effectively solving the data silo problem and strengthening the model's predictive capabilities and the reliability of decision support. The construction of a dynamic feedback mechanism enables real-time response to changes in the construction environment, improving the adaptability and flexibility of investment decisions. Automated modeling capabilities reduce manual intervention, improve model building efficiency, and ensure the feasibility of investment optimization for large-scale power grid projects. The formation of an effective closed-loop feedback mechanism enables dynamic optimization throughout the entire process from data collection and model training to investment decision-making, improving the adaptability and flexibility of investment plans and ensuring the consistency and effectiveness of investment decisions with power grid project implementation.

[0066] Example 2 During peak travel periods like the Spring Festival, summer heatwaves, or extreme weather, existing power distribution networks frequently experience issues such as transformer overload, tripping, and customer complaints. Power grid projects can include power distribution capacity expansion projects.

[0067] In this embodiment, the method further includes: Step Sa: Obtain the second multi-source data for the corresponding area of ​​the project, perform spatiotemporal drift detection processing on the second multi-source data, and obtain the spatiotemporal drift detection results; the second multi-source data includes mobile phone signaling grid data, meteorological monitoring data, power grid operation data, and disaster early warning data.

[0068] In this embodiment, population flow thermal distribution data obtained through operator base stations can be used as mobile signaling grid data. Meteorological monitoring data can be obtained by accessing the meteorological bureau's API, and this data can include real-time data such as temperature, humidity, and wind speed. Power grid operation data can be extracted from the SCADA system, which can include distribution transformer load rate, voltage fluctuations, and fault records. Disaster early warning data can include integrated early warning information for earthquakes, typhoons, and floods (such as real-time data from the China Earthquake Networks Center).

[0069] For example, mobile signaling grid data can be data corresponding to a 500m×500m cellular grid, with data fields including grid ID, timestamp, number of permanent residents, and number of working residents. Meteorological monitoring data can be data collected by an automatic weather station 3km away from the distribution transformer, with data fields including dry-bulb temperature, humidity, wind speed, and typhoon warning level. Power grid operation data can be active power, reactive power, and power factor of the distribution transformer terminal at 15-minute intervals. Disaster early warning data can be typhoon, rainstorm, and urban flooding warnings issued by relevant authorities, and also include latitude and longitude polygon boundaries.

[0070] This allows for the unified alignment of the four types of data that were originally heterogeneous and had different frequencies, reducing drift event location errors and time errors, and providing a highly reliable basis for subsequent load change attribution.

[0071] Step Sb: Based on the spatiotemporal drift detection results, generate a load influence factor matrix; the load influence factor matrix includes population characteristics, meteorological deviation, and disaster attenuation coefficient.

[0072] In this embodiment, the spatiotemporal drift detection result can be quantized into a 3×n matrix (n is the prediction step size). The complete process from drift detection result to load influence factor matrix can include: Step Sb1: Slice the drift events, including the time span and spatial extent: Taking a typhoon drift event as an example, the time span can be from 08:00 on August 1, 2024 to 08:00 on August 4, 2024 (a total of 288 15-minute steps, i.e., n=288). The spatial range is all the grids corresponding to the power supply area of ​​the No. 17 distribution transformer in an old urban area.

[0073] Step Sb2: Calculate the three types of load influence factors to obtain the load influence factor matrix.

[0074] A PostGIS buffer zone is established with the distribution transformer high-voltage pole as the center and R = 1 km. For example, a 250 m raster is placed within the circle with the distribution transformer high-voltage pole as the center and R = 1 km. The population characteristics are obtained by weighting and summing the values ​​using inverse distance squared weights. The population characteristics can be expressed as: Pt=Σ{i∈buffer}pop{i,t}×(1 / di) 2 ) Where P(t) represents population characteristics, and di is the distance (km) from the grid center to the high-voltage pole of the distribution transformer. For example, 12 hours before the typhoon made landfall, Pt surged from 2560 people to 3200 people, and population characteristics = 3200 / 2560 = 1.25.

[0075] Meteorological deviation can be expressed as: WD(t) = Traw(t) - Tbase(t) Where WD(t) represents the meteorological deviation, Traw(t) represents the measured meteorological data of the weather station every 15 minutes, and Tbase(t) represents the average meteorological data of the same period over the past three years. For example, when the meteorological data is temperature, if the measured temperature is 32.4℃ and the average temperature is 28.2℃, then the temperature deviation is 4.2℃.

[0076] The disaster attenuation coefficient can be expressed as: DC(t) = 1 - Poutage(t) Where DC(t) represents the disaster attenuation coefficient, and Poutage(t) represents the expected probability of power outage. Taking a typhoon drift event as an example, the expected probability of power outage can be estimated jointly using the typhoon warning level and a historical fault database. For example, if the typhoon warning level is Level III and the expected probability of power outage is 15%, then the disaster attenuation coefficient is 0.85.

[0077] By using the load impact factor matrix, complex driving forces such as typhoon drift events, weather, and population influx can be explicitly quantified, solving the shortcomings of traditional methods in coupling multidimensional data and avoiding the subjective bias of traditional empirical methods.

[0078] Step Sc: Based on the load influence factor matrix and historical load sequence, a load prediction hyperbola is generated through a pre-trained conditional quantile autoregressive model; the historical load sequence is the active power time series data recorded by power grid equipment at various historical moments, and the load prediction hyperbola includes a median curve and an upper limit curve.

[0079] The historical load series can be active power data at 96 points (every 15 minutes) over the past 365 days. The load influence factor matrix is ​​the spatiotemporal matrix output in step S102. The historical load series and the load influence factor matrix serve as inputs to the conditional quantile autoregressive model.

[0080] In this embodiment, the conditional quantile autoregressive model can be expressed as: Qτ(Yt∣Xt)=β0+∑i=1pβiYt-i+ΓMt(2) Among them, τ=0.5 can generate the median curve, and τ=0.9 can generate the upper limit curve (which can cover 90% of extreme scenarios).

[0081] For example, the load forecast results for a residential area in an old city may include: at 20:00 on August 1, 2024, the corresponding median curve may be 1250 kW, and the corresponding upper limit curve may be 1580 kW; at 21:00 on August 1, 2024, the corresponding median curve may be 1180 kW, and the corresponding upper limit curve may be 1490 kW. The upper limit curve is higher than the median value, which can cover peak scenarios such as concentrated air conditioning startup.

[0082] By simultaneously outputting the median and upper limit curves using the Conditional Quantile Autoregressive (CQAR) model, the coverage of load peaks is improved compared to the traditional mean regression model, reducing the possibility that expansion decisions may fail due to underestimating extreme loads.

[0083] Step Sd: Based on the comparison results of the load forecast hyperbola and the dynamic threshold, determine whether to trigger a power grid project investment decision; the comparison results include the project risk level and the project power shortage parameters. If yes, proceed to step Se.

[0084] In this embodiment, the dynamic threshold can be adjusted in real time according to holidays, temperature, emergencies, etc., reducing the possibility of false alarms or missed alarms due to fixed thresholds in extreme scenarios.

[0085] The instantaneous overload rate can be calculated based on the dynamic threshold and upper limit curve. The instantaneous overload rate can be expressed as: Where ηt represents the instantaneous overload rate, θt represents the upper limit curve, and θt represents the dynamic threshold.

[0086] Based on the comparison between the dynamic threshold and the upper limit curve, the maximum sustained overload time is obtained, which can be expressed as: ΔT = continuously satisfies The duration of the time period (h) Where ΔT represents the maximum continuous overload time.

[0087] The project risk level can be obtained based on the instantaneous overload rate. For example, when ηt < 0, the project risk level is 0, which is a normal situation; when 0 ≤ ηt < 10%, the project risk level is Level I, which is a slight risk situation; when 10% ≤ ηt < 25%, the project risk level is Level II, which is a moderate risk situation; and when ηt ≥ 25%, the project risk level is Level III, which is a severe risk situation.

[0088] The project's power shortage parameters can be expressed as: Where ΔSt represents the project's power shortage parameter, and cosφ can be taken as a typical power factor of 0.9; the result is rounded up to the standard capacity range. For example, ΔSt=max(0,20) / 0.9=22.2kVA, which can be rounded up to 30kVA (minimum standard range).

[0089] In this embodiment, the conditions for triggering the power grid project investment decision can be set as follows: the project risk level is not lower than Level II and the maximum continuous overload time is not less than 2 hours.

[0090] Step Se: Based on a preset economic optimization model, calculate the predicted net present value and internal rate of return corresponding to the power grid project investment; the predicted net present value is a parameter characterizing the absolute return of the power grid project investment, and the internal rate of return is a parameter characterizing the relative return of the power grid project investment.

[0091] In this embodiment, the projected net present value (NPV) can be expressed as: NPV = ∑(future cash flows / (1 + discount rate)^t) - initial investment The internal rate of return (IRR) is the discount rate that makes NPV = 0, representing the expected rate of return of the expansion investment project itself.

[0092] Step Sf: If the predicted net present value is non-negative and the internal rate of return is not lower than the adjusted discount rate, then output an expansion investment instruction; the expansion investment instruction includes the target expansion plan and the required equipment list.

[0093] A non-negative projected net present value (NPV) of ≥0 indicates that, after discounting all future returns back to today using the time value of money, at least the entire investment cost can be recovered. This means the present value of the expansion investment project's returns covers all investment costs and opportunity costs, and even after deducting the cost of capital, it still creates excess value (absolute return meets the target). An IRR ≥ the adjusted discount rate means that the expansion investment project's rate of return is higher than the cost of capital (such as the weighted average cost of capital, WACC), and the risk-adjusted return meets the investor's minimum requirements (relative return meets the target). NPV ≥ 0 ensures the static rationality of the expansion, while IRR ≥ the adjusted discount rate ensures dynamic risk resistance.

[0094] Through a four-layer closed loop encompassing data, models, decision-making, and economics, real-time, economical, and implementable power grid project investment decisions are achieved. Four types of data—mobile phone signaling grid data, meteorological monitoring data, power grid operation data, and disaster early warning data—are automatically aligned within a four-dimensional spatiotemporal grid, reducing drift positioning errors, mitigating abnormal peak values, and achieving the fusion of multi-source heterogeneous data. The conditional quantile autoregressive model outputs median and upper limit curves, providing conservative boundaries, reducing the possibility of underestimating extreme loads, and enabling the generation of a real-time load forecast hyperbola. Based on an economic optimization model, real-time economic optimization calculations for power grid project investment are achieved.

[0095] In this embodiment, the spatiotemporal drift detection processing of the second multi-source data to obtain the spatiotemporal drift detection result includes: The mobile phone signaling grid data, the meteorological monitoring data, the power grid operation data, and the disaster early warning data are uniformly mapped to a four-dimensional spatiotemporal grid to obtain a spatiotemporally aligned second multi-source data cube; For each type of data in the second multi-source data cube, drift detection is performed to obtain an independent drift marker set; based on the preset conflict resolution rules, the authenticity of the independent drift marker set corresponding to each type of data is verified, and the verified joint drift event matrix is ​​obtained. Based on real-time scene markings, the joint drift event matrix is ​​weighted and calculated to obtain a comprehensive spatiotemporal drift score; scene markings include regular day mode, holiday mode, typhoon warning period or regional power outage; Based on the spatiotemporal drift comprehensive score, the spatiotemporal drift detection results are obtained; the spatiotemporal drift detection results include drift type, spatial range and time span, the drift type is a parameter characterizing the cause of distribution transformer load anomalies, the spatial range is a parameter characterizing the distribution transformer area that needs to be expanded, and the time span is a parameter characterizing the urgency of the power grid project.

[0096] Four types of heterogeneous data—mobile phone signaling raster data (spatial resolution ≤ 500m, temporal resolution ≤ 2 hours), meteorological monitoring data (temperature / humidity / wind speed time series), power grid operation data (transformer active power time series), and disaster early warning data (typhoon path / seismic wave attenuation spatial coordinates)—are aligned using the same "space-time-type" coordinate system to form a second multi-source data cube that can be directly processed in parallel. The four-dimensional spatiotemporal grid includes a spatial grid ID, timestamp, data type, and data value.

[0097] A separate drift metric was designed and detected for each data type, reducing both false positive and false negative rates. Specifically: Drift detection of mobile signaling grid data includes: calculating the population difference between grids in adjacent time periods. If the population difference between grids in adjacent time periods is greater than a preset population threshold and the spatial heterogeneity is greater than a preset heterogeneity threshold, it can be marked as a population drift grid. The population difference between grids in adjacent time periods is the amount of population change in the same grid between two adjacent time windows, and the spatial heterogeneity is the degree of unevenness in the population distribution of neighboring grids.

[0098] Drift detection of meteorological monitoring data includes: calculating the meteorological change rate of meteorological data such as temperature or humidity within a preset time period. If the meteorological change rate is greater than a preset meteorological threshold or the typhoon projection distance is less than a preset distance threshold, it can be marked as a meteorological anomaly. The meteorological change rate is the sliding slope of meteorological data such as temperature or humidity within a preset time period, and the typhoon projection distance is the great circle distance from the typhoon center to the high-voltage pole of the distribution transformer.

[0099] Drift detection of power grid operation data includes: calculating load cycle residuals. If the load cycle residuals are greater than a preset residual threshold and the correlation coefficients of adjacent distribution transformers are less than a preset correlation coefficient threshold, it can be marked as abnormal load fluctuation. The load cycle residuals are the difference between the measured load and the load predicted by the SARIMA model, and the correlation coefficients of adjacent distribution transformers are the Pearson correlation coefficients between the target distribution transformer i and its geographically neighboring distribution transformer j.

[0100] Drift detection of disaster early warning data includes: calculating the seismic wave attenuation coefficient; if the seismic wave attenuation coefficient is less than a preset attenuation threshold and the flood impact radius is greater than a preset radius threshold, it can be marked as a disaster-prone area. The flood impact radius can be determined by the radius (km) of the circumcircle of a polygon published by the water resources department.

[0101] The seismic wave attenuation coefficient can be expressed as: α = e^{-0.05t} Where α represents the seismic wave attenuation coefficient, and t is the time (h) from the onset of the disaster. For example, when t = 24h, α = 0.3.

[0102] Based on preset conflict resolution rules, the authenticity of the independent drift marker sets corresponding to each type of data is verified, and the verified joint drift event matrix is ​​output. For example, the preset conflict resolution rules include: if only the power grid load data is abnormal, but there are no abnormalities in meteorological, population, or disaster markers, it can be regarded as an equipment failure, and the corresponding drift can be excluded; if meteorological data anomalies and population drift data coexist, and the spatial overlap is >70%, it can be confirmed as a composite drift event. That is, a sudden increase in power grid load without external factors can be judged as an equipment failure, and drift markers can be excluded to avoid single-source false alarms (such as transformer failures). Meteorological anomalies and population aggregation spatial overlap can be superimposed into composite events, triggering cross-data compensation (such as population migration compensation during cold waves) to enhance event correlation. When data is missing in the disaster-affected area, the interpolation radius R can be automatically expanded and filled with complete neighboring data to ensure spatial continuity.

[0103] Through a pre-set LSTM network, four types of data weights can be dynamically assigned based on real-time scene markings. The regular day mode refers to a calendar date that is neither a statutory holiday nor has any typhoon, rainstorm, or earthquake warning, and there is no planned power outage announcement from the power grid. The holiday mode refers to a calendar date that is a statutory holiday or a long holiday formed by adjusting work schedules (e.g., Spring Festival, National Day, May Day, etc.). The typhoon warning period refers to a typhoon blue or higher warning issued by the meteorological bureau, or a typhoon center less than 200km from the distribution transformer. The regional power outage refers to a planned maintenance or fault power outage announcement issued by the dispatch center, affecting the distribution transformer's power supply area.

[0104] For example, when the scene is marked as a regular day mode, the weight corresponding to the mobile phone signaling grid data can be 0.4, the weight corresponding to the meteorological monitoring data can be 0.3, the weight corresponding to the power grid operation data can be 0.2, and the weight corresponding to the disaster early warning data can be 0.1.

[0105] For each data type, the drift intensity is obtained based on the joint drift event matrix, with a value range of [0,1] (i.e., the normalized drift amplitude). The spatiotemporal drift comprehensive score can be expressed as: S=Σwi×scorei Where S represents the spatiotemporal drift comprehensive score, wi represents the weight of the data, and scorei represents the drift intensity of the data.

[0106] If the spatiotemporal drift comprehensive score is greater than the preset detection threshold, a drift detection result can be generated, which includes drift type, spatial range and time span. Drift type can include data drift, concept drift and composite drift. Spatial range can include drift grid coordinates and disaster impact boundary. Time span can include start timestamp and duration.

[0107] In this embodiment, before determining whether to trigger a power grid project investment decision based on the comparison result of the load forecast hyperbola and the dynamic threshold, the method further includes: Based on the historical load sequence and the historical load forecast sequence, the historical monthly forecast error is calculated, and based on the historical monthly forecast, the basic threshold for monthly updates is obtained; the historical load forecast sequence is the forecast result sequence of the conditional quantile autoregressive model for the load at historical time, and the historical monthly forecast error is the average forecast deviation of the conditional quantile autoregressive model within the monthly cycle. Based on real-time prediction error, meteorological temperature difference, homecoming index, and dummy variables of emergencies, a real-time factor vector is calculated. This real-time factor vector includes an error factor, a holiday factor, and an event factor. The real-time prediction error is the instantaneous deviation at the current moment; the meteorological temperature difference is the difference between the current temperature and the historical average temperature for the same period; the homecoming index is a parameter characterizing population migration within the project's corresponding area; and emergencies include disasters or regional power outages occurring within the project's corresponding area. Based on a trained LSTM network model, a weight vector corresponding to the real-time factor vector is obtained. This weight vector includes the weights corresponding to the error factor, the holiday factor, and the event factor. The dynamic threshold is obtained based on the base threshold, the real-time factor vector, and the weight vector.

[0108] The historical load series (L1) is a time series of active power from distribution transformers (e.g., 96 points / day data). The historical load forecast series (L2) can be generated by a conditional quantile autoregressive model, containing the median and upper limit curve forecasts for each historical time point. The historical monthly forecast error (MAPE) can be calculated as follows: Where N represents the number of monthly data points (e.g., an average of 2,880 15-minute interval points per month).

[0109] In this embodiment, the 95th percentile of historical monthly forecast errors can be extracted as a base threshold. By automatically updating the base threshold at the beginning of each month, it is possible to adapt to seasonal load changes (such as the increase in error caused by air conditioning load in summer).

[0110] The meteorological temperature difference is the difference between the current temperature and the historical average temperature for the same period. The homecoming index is the proportion of the migrant population calculated based on mobile phone signaling grid data (for example, the value can be 0–1). When a disaster / regional power outage event occurs, the dummy variable for the emergency event is 1, otherwise it is 0.

[0111] The calculation process for the error factor can be expressed as follows: aerr=e-|Real-time prediction error-5%| That is, the error is attenuated when it deviates from the reference by 5%, thus suppressing occasional noise.

[0112] The calculation process for the holiday factor can be expressed as follows: aholiday = 1 + 0.6 × Homecoming Index × (1 + 0.2 × Temperature Difference / 10) That is, population migration can be amplified in conjunction with temperature difference; for every 10°C increase in meteorological temperature difference, the sensitivity increases by 20%.

[0113] The calculation process for event factors can be expressed as follows: aevent = 1 + 0.4 × sudden event dummy variable This means that sudden events directly raise the threshold by 40%.

[0114] In this embodiment, the LSTM network model can be trained using historical factor vectors and corresponding historical error sequences to minimize prediction errors and output weight vectors. Backpropagation is used to adjust network parameters and learn the weight patterns for different scenarios. Real-time factor vectors are used as input to the LSTM network model, allowing the pre-trained model to directly output weights. By outputting dynamic weights through the LSTM network model, the false positive rate is reduced, facilitating accurate responses to complex scenarios.

[0115] The formula for synthesizing the dynamic threshold can be expressed as: Dynamic threshold = base threshold × ∑(βk*αk) Where αk represents the real-time factor vector, and βk represents the weight vector corresponding to the real-time factor vector.

[0116] In this embodiment, calculating the predicted net present value and internal rate of return corresponding to the power grid project investment based on a preset economic optimization model includes: Based on the upper limit curve, the current rated capacity of the distribution transformer, and the safety margin coefficient, the capacity expansion requirement is obtained; Based on the aforementioned capacity expansion requirements and the cost of the capacity expansion equipment, obtain the static investment cost; Based on the project's power shortage parameters and the regional power outage value of the corresponding area, the expected load loss cost is obtained. Based on the project risk level, adjust the initial benchmark discount rate to obtain the target benchmark discount rate; Based on the static investment cost, the expected cost of loss of load, and the target benchmark discount rate, calculate the projected net present value and the internal rate of return.

[0117] In this embodiment, the safety margin coefficient can be obtained based on the ratio of load rate to current rated capacity. For example, when the load rate > 95% * current rated capacity, the safety margin coefficient can be 0.75, which is a high-risk area; when 85% * current rated capacity ≤ load rate ≤ 95% * current rated capacity, the safety margin coefficient can be 0.8, which is a medium-risk area; when the load rate < 85% * current rated capacity, the safety margin coefficient can be 0.85. The capacity expansion requirement can be expressed as: Capacity expansion requirement = ceil(max(upper limit curve) / safety margin coefficient) In this embodiment, the cost of the capacity expansion equipment may include the basic cost and the copper price. The real-time copper price (PCu) can be updated every 30 minutes by connecting to the Shanghai Futures Exchange (SHFE) via API.

[0118] The formula for calculating static investment cost can be expressed as: Cstatic=Snew×(PCu×1.05+Cbase) Wherein, Cstatic represents static investment cost, Snew represents capacity expansion demand, PCu represents real-time copper price, and Cbase represents basic construction cost (including iron core and insulation materials).

[0119] In this embodiment, the electronic device stores a mapping relationship between the area type of the project's corresponding area and the area power outage value. For example, when the area type is an industrial area, the area power outage value can be 15.0 yuan / kWh; when the area type is a residential area, the area power outage value can be 4.5 yuan / kWh.

[0120] The formula for calculating the expected loss of load can be expressed as: Where Δt = 1 hour, the cumulative power supply loss over 90 days can be calculated.

[0121] The electronic device also stores the mapping relationship between the project risk level and the adjustment amount of the initial benchmark discount rate. For example, when the project risk level is high risk, the target benchmark discount rate = initial benchmark discount rate - 2%. In this embodiment, the calculation of the predicted net present value and internal rate of return is prior art and will not be described in detail here.

[0122] In this embodiment, after the output expansion investment instruction, the following is also included: If the current power grid project is successfully put into operation, the operation data after the power grid project is collected in real time; the operation data includes the privacy-preserving actual load rate, actual investment amount, and actual net present value; the privacy is achieved by applying differential privacy noise to the operation data; Based on the operational data, the expansion success rate index is calculated, and based on the expansion success rate index, the project level corresponding to the current power grid project is obtained. Based on the project level, the initial benchmark discount rate for subsequent power distribution capacity expansion investment projects will be dynamically adjusted.

[0123] In this embodiment, the actual load rate can be transmitted back in real time by the smart meter, the actual investment amount can be obtained by connecting to the financial system, the actual investment amount includes equipment and construction costs, and the actual net present value can be calculated retrospectively by the economic model.

[0124] The success rate index for capacity expansion can be expressed as: Expansion success rate index = First weight * (Actual load rate after expansion / Design load rate) + Second weight * (Actual net present value / Investment amount) The electronic device stores a mapping relationship between a first weight, a second weight, and a region type. For example, when the region type is an industrial zone, the first weight can be 0.6 and the second weight can be 0.4.

[0125] The electronic device stores the mapping relationship between the expansion success rate index and the project level, and between the project level and the adjustment amount of the initial benchmark discount rate. For example, when the expansion success rate index corresponds to the interval [0.85, 1.0], the project level is high efficiency, and the adjusted initial benchmark discount rate = the current initial benchmark discount rate - 3%.

[0126] In this embodiment, before the output expansion investment instruction, the following is also included: Based on the long-term planning data of the corresponding area of ​​the project, the current capacity expansion plan is validated under a scenario; the long-term planning data includes spatial planning data, industrial relocation plan data, and distributed energy access planning data. If the scenario verification result indicates that the capacity expansion requirement corresponding to the current capacity expansion plan is redundant within the long-term period, then a capacity tiered decision-making mechanism is initiated. The capacity tiered decision-making mechanism includes: using the difference between the medium-term load forecast curve and the long-term load forecast curve as the redundant capacity, and using the redundant capacity as a constraint to re-optimize the current capacity expansion plan and generate a tiered capacity expansion path. The tiered capacity expansion path includes the initial capacity expansion, medium-term scalable interfaces, and long-term reserved capacity. The medium-term load forecast curve and the long-term load forecast curve are obtained by the conditional quantile autoregressive model using the spatial planning data as a new exogenous variable and re-rolling the forecast, with the forecast step size corresponding to the medium-term planning period and the long-term planning period, respectively. Based on the tiered capacity expansion path, update the static investment cost, the expected offload cost, and the target benchmark discount rate, and recalculate the predicted net present value and the internal rate of return. If the updated predicted net present value is still non-negative and the internal rate of return is not lower than the adjusted discount rate, then the target capacity expansion plan is generated based on the current capacity expansion plan and the tiered capacity expansion path.

[0127] Before formally issuing the "expansion investment order," it's essential to examine long-term planning data to determine whether the currently calculated capacity expansion will become redundant in the next 5-10 years. Long-term planning data can include: land use planning (i.e., whether the land has been designated for residential, industrial, or green space); industrial relocation plans (i.e., whether large data centers, shopping malls, or factories will be built); and distributed energy access plans (i.e., whether rooftop solar power and energy storage will offset some of the load locally).

[0128] By overlaying various planning data, a simulated future load is generated and compared with the calculated capacity expansion demand. If the future load is less than or equal to the expanded capacity, there is redundancy in the distribution transformer capacity, resulting in idle assets. If the future load is greater than the expanded capacity, there is no redundancy in the distribution transformer capacity, and the original plan can be implemented directly.

[0129] The medium-term load forecast curve is the load growth curve projected 3 to 5 years from now, used as the basis for the initial capacity expansion. The long-term load forecast curve is the load growth curve projected 8 to 10 years from now, representing the final saturation load. Redundancy capacity is the difference between the medium-term and long-term load forecast curves, representing the capacity built in the initial phase but ultimately not used.

[0130] The tiered capacity expansion path doesn't involve building all the redundant capacity at once. Instead, it divides capacity expansion into three phases: initial, medium, and long-term. The initial capacity expansion only meets medium-term load requirements. The medium-term expansion includes expandable interfaces, with busbars, foundations, and corridors reserved in the initial phase to avoid repeated excavation in the future. The long-term capacity is reserved by adding modules to the interfaces, allowing for a smooth upgrade to the planned capacity.

[0131] Since the tiered capacity expansion path differs from the current capacity expansion plan, the static investment cost, expected loss of load cost, and target benchmark discount rate may all change. Therefore, the projected net present value and internal rate of return can be recalculated based on the new capacity expansion plan. The specific calculation method is the same as the aforementioned method and will not be elaborated here.

[0132] When the predicted net present value and internal rate of return of the new capacity expansion plan (i.e., the plan formed by the tiered capacity expansion path of the current capacity expansion plan) meet the conditions, the new capacity expansion plan will be used as the target capacity expansion plan to guide the investment of power grid projects.

[0133] In this embodiment, the learning method of the conditional quantile autoregressive model includes: For each regional node, the second multi-source data is acquired locally at that regional node, and a conditional quantile autoregressive sub-model is independently trained based on the second multi-source data to obtain local parameters; the regional node is a power distribution business unit with independent data and computing capabilities. For each region node, differential privacy noise is applied to the local parameters corresponding to the region node to generate privacy parameters; for each region node, the privacy parameters corresponding to the region node are packaged with the load quantile prediction results of the local validation set into a zero-knowledge proof. The coordinating node verifies the zero-knowledge proof submitted by each region node one by one; if the current region node passes the verification, it enters the aggregation phase; if the current region node fails the verification, it is retrained. The coordinating node performs a weighted federated average on the verified privacy parameters, with the weight being the reciprocal of the average quantile loss of the verification set of each regional node, to obtain the global parameters; Using the joint validation set resulting from the merging of the local validation sets of all validated regional nodes as a benchmark, the quantile loss volatility of the joint validation set after k consecutive rounds of aggregation is calculated. If the quantile loss volatility is less than a preset threshold after k consecutive rounds of aggregation, the global parameters are used as the final convergent model and training is terminated, resulting in the trained conditional quantile autoregressive model.

[0134] Each regional node (such as the provincial power grid dispatch center) independently trains a conditional quantile autoregressive sub-model based on local second multi-source data (load sequence, meteorological data, population migration index).

[0135] Add Laplace noise ∈ ~Laplace(0,Δf / ∈) to the local parameters, where the privacy budget ∈ = 0.8 and the sensitivity Δf = 1.5 (set according to the parameter range). Output the privacy parameter θpriv = local parameter + ∈.

[0136] In this embodiment, the proof includes: the regional node uses privacy parameters to predict the local validation set load quantile. A ZKP evidence package is generated: ZKP-Package = {privacy parameters, local validation set load quantile, hash value}, where the hash value is used for result verifiability.

[0137] The quantile loss threshold can be dynamically adjusted according to regional load characteristics (e.g., 7% for industrial areas and 10% for residential areas). Failed nodes are learned incrementally (based on the parameter initialization of the previous round) to accelerate convergence.

[0138] Based on the same technical concept, this application also provides an investment optimization calculation device based on a multimodal artificial intelligence model, such as... Figure 6 As shown, the investment optimization calculation device 200 based on a multimodal artificial intelligence model mainly includes: The data acquisition module 201 is used to acquire first multi-source data of the power grid project. The first multi-source data includes structured data, unstructured data, and environmental parameters. The structured data includes at least the planned start date, planned commissioning date, planned total investment, total investment approved in the feasibility study, approval date, line length, and substation capacity. The unstructured data includes at least the feasibility study report text, construction service contract text, material procurement contract text, construction progress images, GIS topology, and DEM elevation data. The environmental parameters include at least the historical and predicted meteorological time series and geographical location topology data of the construction area. The format processing module 202 is used to perform unified format processing on the integrated first multi-source data to form a multimodal feature vector that can be directly input into the pre-trained model. The model prediction module 203 is used to input the multimodal feature vectors into the trained multimodal construction period calculation model and the monthly investment amount calculation model respectively, so as to obtain the predicted values ​​of the construction period, settlement period, and final account period of the power grid project, as well as the monthly investment ratio sequence. The deviation calculation module 204 is used to collect progress images and the latest meteorological data of the construction site in real time, calculate the deviation between the actual progress and the predicted progress, and incrementally update the construction period calculation model and the monthly investment amount calculation model through the continuous learning interface, and adjust the investment ratio of the remaining months in real time. The reallocation module 205 is used to reallocate annual investment amounts among multiple power grid projects based on the updated monthly investment ratio and according to the capital utilization threshold, so as to achieve cross-power grid project capital balance. The decision output module 206 is used to generate the annual investment amount, milestone plan and capital expenditure curve of a single power grid project based on the adjusted annual investment plan, and serves as the final output of the power grid company's intelligent investment decision-making.

[0139] In one example, the module in any of the above devices may be one or more integrated circuits configured to implement the above methods, such as one or more application-specific integrated circuits (ASICs), or one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs), or a combination of at least two of these integrated circuit forms.

[0140] For example, when modules in a device can be implemented via a processing element scheduler, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor capable of calling programs. Alternatively, these modules can be integrated together as a system-on-a-chip (SOC).

[0141] In this application, various objects such as messages / information / devices / network elements / systems / apparatus / actions / operations / processes / concepts may be named. It is understood that these specific names do not constitute a limitation on the relevant objects. The names may be changed depending on the scenario, context, or usage habits. The understanding of the technical meaning of the technical terms in this application should be mainly determined from their functions and technical effects embodied / performed in the technical solution.

[0142] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0143] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0144] Based on the same technical concept, this application also provides an electronic device, such as... Figure 7 As shown, the electronic device 300 includes a processor 301 and a memory 302, and may further include one or more of an information input / output (I / O) interface 303, a communication component 304, and a communication bus 305.

[0145] The processor 301 controls the overall operation of the electronic device 300 to complete all or part of the steps in the investment optimization calculation method based on the multimodal artificial intelligence model described above. The memory 302 stores various types of data to support the operation of the electronic device 300. This data may include, for example, instructions for any application or method operating on the electronic device 300, as well as application-related data. The memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as one or more of Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0146] I / O interface 303 provides an interface between processor 301 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 304 is used to test wired or wireless communication between electronic device 300 and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 304 may include a Wi-Fi component, a Bluetooth component, and an NFC component.

[0147] The communication bus 305 may include a path for transmitting information between the aforementioned components. The communication bus 305 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 305 can be divided into an address bus, a data bus, a control bus, etc.

[0148] The electronic device 300 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the investment optimization calculation method based on the multimodal artificial intelligence model given in the above embodiments.

[0149] Electronic device 300 may include, but is not limited to, mobile terminals such as digital broadcast receivers, PDAs (personal digital assistants), and PMPs (portable multimedia players), as well as fixed terminals such as digital TVs and desktop computers, and may also be servers.

[0150] Based on the same technical concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described xxx method.

[0151] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] The terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0153] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0154] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0155] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. An investment optimization calculation method based on a multimodal artificial intelligence model, characterized in that, include: Acquire the first multi-source data of the power grid project, which includes structured data, unstructured data and environmental parameters; The structured data includes at least the planned start date, planned commissioning date, planned total investment, total investment approved in the feasibility study, approval date, line length, and substation capacity; the unstructured data includes at least the feasibility study report text, construction service contract text, material procurement contract text, construction progress images, GIS topology and DEM elevation data; the environmental parameters include at least the historical and predicted meteorological time series and geographical location topology data of the construction area. The integrated first multi-source data is uniformly formatted to form a multimodal feature vector that can be directly input into the pre-trained model; The multimodal feature vectors are input into the trained multimodal construction period calculation model and monthly investment amount calculation model respectively to obtain the predicted values ​​of the construction period, settlement period, and final account period of the power grid project, as well as the monthly investment ratio sequence. The system collects progress images and the latest meteorological data from the construction site in real time, calculates the deviation between the actual progress and the predicted progress, and incrementally updates the construction period calculation model and the monthly investment amount calculation model through a continuous learning interface, and adjusts the investment ratio for the remaining months in real time. Based on the updated monthly investment ratio, the annual investment amount is redistributed among multiple power grid projects according to the capital utilization rate threshold in order to achieve cross-power grid project capital balance. Based on the adjusted annual investment plan, the annual investment amount, milestone plan and capital expenditure curve of a single power grid project are generated, and serve as the final output of the power grid company's intelligent investment decision-making. The method further includes: Acquire second multi-source data corresponding to the project area, perform spatiotemporal drift detection processing on the second multi-source data, and obtain spatiotemporal drift detection results; the second multi-source data includes mobile phone signaling grid data, meteorological monitoring data, power grid operation data, and disaster early warning data; the process of performing spatiotemporal drift detection processing on the second multi-source data to obtain spatiotemporal drift detection results includes: uniformly mapping the mobile phone signaling grid data, the meteorological monitoring data, the power grid operation data, and the disaster early warning data to a four-dimensional spatiotemporal grid to obtain a spatiotemporally aligned second multi-source data cube; performing drift detection on each type of data in the second multi-source data cube to obtain an independent drift marker set; based on a preset... The conflict resolution rules verify the authenticity of the independent drift marker sets corresponding to each type of data and obtain the verified joint drift event matrix. Based on real-time scene markers, the joint drift event matrix is ​​weighted and calculated to obtain a comprehensive spatiotemporal drift score. Scene markers include regular daily patterns, holiday patterns, typhoon warning periods, or regional power outages. Based on the comprehensive spatiotemporal drift score, the spatiotemporal drift detection results are obtained. The spatiotemporal drift detection results include drift type, spatial range, and time span. The drift type is a parameter characterizing the cause of abnormal distribution transformer load, the spatial range is a parameter characterizing the distribution transformer area that needs expansion, and the time span is a parameter characterizing the urgency of the power grid project. Based on the spatiotemporal drift detection results, a load impact factor matrix is ​​generated; the load impact factor matrix includes population characteristics, meteorological deviation, and disaster attenuation coefficient. Based on the load influence factor matrix and historical load sequence, a load prediction hyperbola is generated through a pre-trained conditional quantile autoregressive model; the historical load sequence is the active power time series data recorded by power grid equipment at various historical moments, and the load prediction hyperbola includes a median curve and an upper limit curve. Based on the comparison results of the load forecast hyperbola and the dynamic threshold, it is determined whether to trigger a power grid project investment decision; the comparison results include the project risk level and the project power shortage parameters. If so, then based on the preset economic optimization model, the predicted net present value and internal rate of return corresponding to the power grid project investment are calculated; the predicted net present value is a parameter characterizing the absolute return of the power grid project investment, and the internal rate of return is a parameter characterizing the relative return of the power grid project investment. If the predicted net present value is non-negative and the internal rate of return is not lower than the adjusted discount rate, then an expansion investment instruction is output; the expansion investment instruction includes the target expansion plan and the required equipment list.

2. The method according to claim 1, characterized in that, The training methods for the multimodal construction period calculation model and the monthly investment amount calculation model include: The merged structured data, unstructured data, and time series data are used to construct a unified dataset; Using AutoGluon's MultiModalPredictor, Transformer, CNN, and time series models are automatically searched and integrated on the unified dataset to perform multi-label regression training on construction period, settlement period, and final account period, resulting in a multimodal period calculation model. Based on the multimodal construction period calculation model, using AutoGluon's TimeSeriesPredictor, SeasonalNaive, ETS, DeepAR, and PatchTST time series models are automatically combined. With the construction period prediction value, static characteristics of the power grid project, historical investment, construction progress, meteorological information, and DEM elevation data as inputs, a monthly investment ratio sequence is generated to obtain a monthly investment amount calculation model. Based on cross-validation and automatic hyperparameter tuning, the model structure and weights of the multimodal construction period calculation model and the monthly investment amount calculation model are determined, and the model accuracy is verified on the test set. Based on the continuous learning framework, construction deviation data is received in real time, and the multimodal construction period calculation model and the monthly investment amount calculation model are incrementally updated.

3. The method according to claim 2, characterized in that, The unified formatting process includes: The structured data is subjected to numerical processing, null padding, multicollinearity removal, and logarithmic standardization. Based on a pre-defined large language model, the necessity of construction, construction scale, construction requirements, and material supply requirements are extracted from the unstructured data and tagged accordingly. Based on a preset visual model, key equipment and construction stages are identified from the construction progress images and output with semantic labels. For the GIS topology and DEM data, extract the terrain type, path length and terrain complexity of the construction path and generate terrain feature vectors; For meteorological time series, extract the number of extreme weather days and generate the number of days of work stoppage due to weather impact; Based on a unified CSV or tensor format, the processed first multi-source data is aligned and serialized to form the multimodal feature vector that can be directly input into the pre-trained model.

4. The method according to claim 1, characterized in that, Before determining whether to trigger a power grid project investment decision based on the comparison results of the load forecast hyperbola and the dynamic threshold, the following steps are also included: Based on the historical load sequence and the historical load forecast sequence, the historical monthly forecast error is calculated, and based on the historical monthly forecast, the basic threshold for monthly updates is obtained; the historical load forecast sequence is the forecast result sequence of the conditional quantile autoregressive model for the load at historical time, and the historical monthly forecast error is the average forecast deviation of the conditional quantile autoregressive model within the monthly cycle. Based on real-time prediction error, meteorological temperature difference, homecoming index, and dummy variables of emergencies, a real-time factor vector is calculated. The real-time factor vector includes error factor, holiday factor, and event factor. The real-time prediction error is the instantaneous deviation corresponding to the current moment. The meteorological temperature difference is the difference between the current temperature and the historical average temperature for the same period. The homecoming index is a parameter characterizing the population migration situation in the area corresponding to the project. Emergencies include disasters or regional power outages occurring in the area corresponding to the project. Based on the trained LSTM network model, obtain the weight vector corresponding to the real-time factor vector; the weight vector includes the weights corresponding to the error factor, the weights corresponding to the holiday factor, and the weights corresponding to the event factor; The dynamic threshold is obtained based on the base threshold, the real-time factor vector, and the weight vector.

5. The method according to claim 4, characterized in that, Following the output expansion investment instruction, the following is also included: If the current power grid project is successfully put into operation, the operation data after the power grid project is collected in real time; the operation data includes the privacy-preserving actual load rate, actual investment amount, and actual net present value; the privacy is achieved by applying differential privacy noise to the operation data; Based on the operational data, the expansion success rate index is calculated, and based on the expansion success rate index, the project level corresponding to the current power grid project is obtained. Based on the project level, the initial benchmark discount rate for subsequent power distribution capacity expansion investment projects will be dynamically adjusted.

6. The method according to claim 4, characterized in that, Prior to the output expansion investment instruction, the following is also included: Based on the long-term planning data of the corresponding area of ​​the project, the current capacity expansion plan is validated under a scenario; the long-term planning data includes spatial planning data, industrial relocation plan data, and distributed energy access planning data. If the scenario verification result indicates that the capacity expansion requirement corresponding to the current capacity expansion plan is redundant within the long-term period, then a capacity tiered decision-making mechanism is initiated. The capacity tiered decision-making mechanism includes: using the difference between the medium-term load forecast curve and the long-term load forecast curve as the redundant capacity, and using the redundant capacity as a constraint to re-optimize the current capacity expansion plan and generate a tiered capacity expansion path. The tiered capacity expansion path includes the initial capacity expansion, medium-term scalable interfaces, and long-term reserved capacity. The medium-term load forecast curve and the long-term load forecast curve are obtained by the conditional quantile autoregressive model using the spatial planning data as a new exogenous variable and re-rolling the forecast, with the forecast step size corresponding to the medium-term planning period and the long-term planning period, respectively. Based on the tiered capacity expansion path, update the static investment cost, expected unloaded cost, and target benchmark discount rate, and recalculate the predicted net present value and the internal rate of return. If the updated predicted net present value is still non-negative and the internal rate of return is not lower than the adjusted discount rate, then the target capacity expansion plan is generated based on the current capacity expansion plan and the tiered capacity expansion path.

7. An investment optimization calculation device based on a multimodal artificial intelligence model applied to the method described in any one of claims 1 to 6, characterized in that, include, The data acquisition module is used to acquire the first multi-source data of the power grid project. The first multi-source data includes structured data, unstructured data, and environmental parameters. The structured data includes at least the planned start date, planned commissioning date, planned total investment, total investment approved in the feasibility study, approval date, line length, and substation capacity. The unstructured data includes at least the feasibility study report text, construction service contract text, material procurement contract text, construction progress images, GIS topology, and DEM elevation data. The environmental parameters include at least the historical and predicted meteorological time series and geographical location topology data of the construction area. The format processing module is used to perform unified formatting processing on the integrated first multi-source data to form a multimodal feature vector that can be directly input into the pre-trained model; The model prediction module is used to input the multimodal feature vectors into the trained multimodal construction period calculation model and the monthly investment amount calculation model respectively, so as to obtain the predicted values ​​of the construction period, settlement period, and final account period of the power grid project, as well as the monthly investment ratio sequence. The deviation calculation module is used to collect progress images and the latest meteorological data at the construction site in real time, calculate the deviation between the actual progress and the predicted progress, and incrementally update the construction period calculation model and the monthly investment amount calculation model through the continuous learning interface, and adjust the investment ratio of the remaining months in real time. The reallocation module is used to reallocate annual investment amounts among multiple power grid projects based on the updated monthly investment ratio and according to the capital utilization threshold, so as to achieve cross-power grid project capital balance. The decision output module is used to generate the annual investment amount, milestone plan and capital expenditure curve of a single power grid project based on the adjusted annual investment plan, and serves as the final output of the power grid company's intelligent investment decision-making.

8. An electronic device, characterized in that, It includes a processor and a memory, wherein the processor is coupled to the memory; The processor is configured to execute a computer program stored in the memory, causing the electronic device to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It includes a computer program or instructions that, when run on a computer, cause the computer to perform the method as described in any one of claims 1 to 6.

Images