A power grid material demand prediction method and system

Abstract

The application discloses a kind of power grid material demand prediction method and system, applied to material prediction technical field, method includes obtaining the engineering construction data of target power grid project, material supply plan data and historical supply data;Based on engineering construction data and material supply plan data, the target power grid project is divided and handled, obtains several project stages and the characteristic data corresponding to several project stages;Several project stages and the characteristic data corresponding to several project stages are dynamically constrained and carded, and material supply time window is obtained;Historical supply data and material supply time window are input into the demand prediction model built and handled, and material demand prediction value is obtained.The method provided by the application can solve the technical problem that the prediction result deviates greatly from the actual demand in the prior art, realizes phased dynamic prediction based on the gradual evolution characteristics of project information, and effectively reduces the risk of material accumulation and supply shortage.

Description

Technical Field

[0001] This invention relates to the field of material forecasting technology, and in particular to a method and system for forecasting power grid material demand. Background Technology

[0002] Power grid construction is characterized by a long overall cycle, information asymmetry at each stage, and strong coupling between material demand and project progress. From power grid planning to material bidding, project information exhibits a gradual evolution from coarse to refined and from vague to definite.

[0003] In existing technologies, material demand forecasting methods are generally based on the quota ratio of the project budget. This method decomposes the total investment of the project budget according to the pre-set quota ratio of equipment costs, material costs, etc., and then combines the unit price quota of various materials to back-calculate the required quantity of each category of materials, forming a material demand list. However, this method is difficult to capture the strong dynamic relationship between material demand and project progress, which will lead to a large deviation between the forecast results and the actual demand, resulting in the accumulation of materials in advance or the shortage of supply later, which in turn leads to the consequences of idle work and material shortage or increased capital occupation costs. Summary of the Invention

[0004] This invention provides a method and system for forecasting power grid material demand to solve the technical problem of large deviation between forecast results and actual demand in the prior art. It realizes phased dynamic forecasting based on the progressive evolution characteristics of project information, so that the material demand forecast results are strongly coupled with the actual progress of the project, significantly improving the forecast accuracy and effectively reducing the risk of material backlog and supply shortage.

[0005] To address the aforementioned technical problems, this invention provides a method for forecasting power grid material demand, the method comprising: Obtain engineering construction data, material supply plan data, and historical supply data for the target power grid project; Based on the engineering construction data and the material supply plan data, the target power grid project is divided into several project stages and characteristic data corresponding to the project stages. Dynamic constraints are applied to several project phases and the corresponding feature data to obtain the material supply time window. The historical supply data and the material supply time window are input into the constructed demand forecasting model for processing to obtain the material demand forecast value. The construction process is configured to use a two-layer stacked long short-term memory neural network structure, with random deactivation layers added between the layers. The historical supply data and the material supply time window are processed, and the material demand forecast value is output through a fully connected layer.

[0006] As one preferred embodiment, the process of dividing the target power grid project into several project phases and corresponding feature data for each project phase includes: Obtain the project time series from the engineering construction data and the planned procurement time nodes from the material supply plan data; Based on the project timeline and the planned procurement timeline, the target power grid project is divided into several project phases, including the pre-feasibility study phase, the feasibility study-preliminary design phase, and the post-preliminary design phase. For each project phase, engineering attribute parameters corresponding to that phase are extracted from the engineering construction data as feature data for that phase; wherein, the engineering attribute parameters include at least the planned investment amount, planned construction period, and material category list for that phase.

[0007] As one preferred embodiment, the step of dynamically constraining and organizing the characteristic data corresponding to several project stages to obtain the material supply time window includes: Obtain the planned start time and preliminary design approval time from the feature data; The latest delivery time is determined based on the planned start time, and the earliest delivery time is determined based on the preliminary approval time and the preset maximum supply cycle. Based on the latest arrival time and the earliest arrival time, a material supply time window is generated, which includes at least the minimum demand value and the maximum demand value.

[0008] As one preferred embodiment, the demand prediction model further includes a feature attention gate and a scale fusion gate; The step of inputting the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast includes: Obtain feature data corresponding to the aforementioned project stages, and input the feature data into the feature attention gate to obtain weighted features; The weighted features, the historical supply data, and the material supply time window are input into the two-layer stacked long short-term memory neural network. After processing by the random deactivation layer, the multi-scale hidden layer output is obtained. The output of the multi-scale hidden layer is input to the scale fusion gate for nonlinear fusion, and the predicted value of material demand is output through the fully connected layer.

[0009] As one preferred embodiment, after obtaining the predicted material demand, the power grid material demand forecasting method further includes: Obtain the predicted material demand and the actual delivery data of the target power grid project; Based on the predicted demand for materials and the actual delivery data, the prediction deviation rate is determined; When the prediction deviation rate exceeds a preset threshold, the actual delivery data is used as a new training sample to incrementally learn the demand prediction model, update the network parameters of the demand prediction model, and obtain an optimized demand prediction model.

[0010] The present invention also provides a power grid material demand forecasting system, comprising: The acquisition module is used to acquire engineering construction data, material supply plan data, and historical supply data of the target power grid project; The segmentation module is used to segment the target power grid project based on the engineering construction data and the material supply plan data, to obtain several project stages and feature data corresponding to the project stages. The constraint module is used to dynamically constrain and sort out the characteristic data corresponding to the project stages to obtain the material supply time window. The processing module is used to input the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast value. The construction process is configured to use a two-layer stacked long short-term memory neural network structure, with random deactivation layers added between the layers, to process the historical supply data and the material supply time window, and output the material demand forecast value through a fully connected layer.

[0011] As one preferred embodiment, the process of dividing the target power grid project into several project phases and corresponding feature data for each project phase includes: Obtain the project time series from the engineering construction data and the planned procurement time nodes from the material supply plan data; Based on the project timeline and the planned procurement timeline, the target power grid project is divided into several project phases, including the pre-feasibility study phase, the feasibility study-preliminary design phase, and the post-preliminary design phase. For each project phase, engineering attribute parameters corresponding to that phase are extracted from the engineering construction data as feature data for that phase; wherein, the engineering attribute parameters include at least the planned investment amount, planned construction period, and material category list for that phase.

[0012] As one preferred embodiment, the step of dynamically constraining and organizing the characteristic data corresponding to several project stages to obtain the material supply time window includes: Obtain the planned start time and preliminary design approval time from the feature data; The latest delivery time is determined based on the planned start time, and the earliest delivery time is determined based on the preliminary approval time and the preset maximum supply cycle. Based on the latest arrival time and the earliest arrival time, a material supply time window is generated, which includes at least the minimum demand value and the maximum demand value.

[0013] As one preferred embodiment, the demand prediction model further includes a feature attention gate and a scale fusion gate; The step of inputting the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast includes: Obtain feature data corresponding to the aforementioned project stages, and input the feature data into the feature attention gate to obtain weighted features; The weighted features, the historical supply data, and the material supply time window are input into the two-layer stacked long short-term memory neural network. After processing by the random deactivation layer, the multi-scale hidden layer output is obtained. The output of the multi-scale hidden layer is input to the scale fusion gate for nonlinear fusion, and the predicted value of material demand is output through the fully connected layer.

[0014] As one preferred embodiment, after obtaining the predicted material demand, the power grid material demand forecasting system further includes: Obtain the predicted material demand and the actual delivery data of the target power grid project; Based on the predicted demand for materials and the actual delivery data, the prediction deviation rate is determined; When the prediction deviation rate exceeds a preset threshold, the actual delivery data is used as a new training sample to incrementally learn the demand prediction model, update the network parameters of the demand prediction model, and obtain an optimized demand prediction model.

[0015] Compared with the prior art, the beneficial effects of the present invention are at least one of the following: This invention acquires engineering construction data, material supply plan data, and historical supply data of a target power grid project; based on the engineering construction data and the material supply plan data, the target power grid project is divided into several project stages and corresponding feature data; the project stages and their corresponding feature data are dynamically constrained to obtain a material supply time window; the historical supply data and the material supply time window are input into a pre-constructed demand forecasting model for processing to obtain a predicted material demand value. The construction process is configured to use a two-layer stacked long short-term memory neural network structure, with randomly deactivated layers added between layers. The historical supply data and the material supply time window are processed, and the predicted material demand value is output through a fully connected layer.

[0016] Compared with existing technologies, this invention acquires engineering construction data, material supply plan data, and historical supply data of the target power grid project. First, based on the engineering construction data and material supply plan data, the project is divided into several project stages and corresponding characteristic data for each stage. This ensures that the predicted progress matches the gradual evolution of project information from coarse to fine, solving the problem that the static breakdown of quota ratios in existing technologies cannot adapt to the dynamic evolution of the project. Next, dynamic constraints are applied to each project stage and its characteristic data to obtain material supply time windows, explicitly establishing the temporal coupling relationship between material demand and project progress. Finally, historical supply data and material supply time windows are input into a demand prediction model constructed using a two-layer stacked long short-term memory neural network structure. This model prevents overfitting and improves generalization ability through inter-layer random deactivation layers, and outputs predicted material demand values ​​through fully connected layers. This further improves prediction accuracy based on stage alignment and temporal coupling constraints. The synergistic effect of these three elements deeply couples the prediction results with the actual project progress, effectively reducing prediction deviations and avoiding material backlogs or supply shortages. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a power grid material demand forecasting method in one embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a power grid material demand forecasting system in one embodiment of the present invention; Figure label: Among them, 11 is the acquisition module; 12 is the partitioning module; 13 is the constraint module; and 14 is the processing module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] One embodiment of the present invention provides a method for forecasting power grid material demand. For details, please refer to [link / reference]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating a power grid material demand forecasting method according to one embodiment of the present invention. The method includes: S1: Obtain engineering construction data, material supply plan data, and historical supply data for the target power grid project; S2: Based on the engineering construction data and the material supply plan data, the target power grid project is divided into several project stages and feature data corresponding to the project stages. S3: Dynamically constrain and sort out the feature data corresponding to the project stages to obtain the material supply time window; S4: Input the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast value. The construction process is configured to use a two-layer stacked long short-term memory neural network structure, with random deactivation layers added between layers, to process the historical supply data and the material supply time window, and output the material demand forecast value through a fully connected layer.

[0021] Specifically, this involves acquiring engineering construction data, material supply plan data, and historical supply data for the target power grid project.

[0022] The engineering construction data is obtained from the power grid company's project management system and design management system, specifically including bill of quantities and budget information extracted from feasibility study reports, preliminary designs and construction drawings.

[0023] Material supply plan data is obtained from the ERP system or material procurement management platform of the material management department. Specifically, it includes the prepared annual, quarterly and monthly material demand plans, material bidding and procurement batch arrangements, and the planned arrival time and inventory turnover data of various materials.

[0024] Historical supply data is obtained from the material management archives and procurement system history records of completed power grid projects. Specifically, it includes the actual types, quantities, unit prices, delivery times, and consumption records of materials purchased for similar projects, as well as comparative data between historical project budgets and actual material expenditures.

[0025] The above data was cleaned, desensitized, and standardized before further processing.

[0026] The target power grid project is divided into several project stages and corresponding feature data, including: obtaining the project time series from the engineering construction data and the planned procurement time nodes from the material supply plan data; dividing the target power grid project into several project stages based on the project time series and the planned procurement time nodes, wherein the project stages include the pre-feasibility study stage, the feasibility study-preliminary design stage, and the post-preliminary design stage; for each project stage, extracting the corresponding engineering attribute parameters from the engineering construction data as feature data for that stage; wherein the engineering attribute parameters include at least the planned investment amount, planned construction period, and material category list for that stage.

[0027] In a specific embodiment, in view of the differences in information completeness at different stages of power grid engineering, this invention divides the project pool into three independent feature subspaces for the first time: the pre-feasibility study stage, the feasibility study-preliminary design stage, and the post-preliminary design stage. Features of different modules are extracted and assigned dynamic weights to each subspace.

[0028] Pre-feasibility study stage: only macro information such as construction scale, voltage level, and planned start time is included, and statistical characteristics such as regional construction intensity and historical material consumption coefficient of similar projects are extracted, with the weight coefficient initialized to 0.2.

[0029] Feasibility study - preliminary design stage: includes information such as investment amount and construction capacity, extracts features such as historical unit material consumption, and initializes the weight coefficient to 0.3.

[0030] Post-initial design stage: Includes information such as construction plan and start date, extracts features such as milestone node plan, and initializes the weight coefficient to 0.5.

[0031] The weighting coefficients are dynamically updated as the project progresses. When the project moves from one phase to the next, the weighting adjustment mechanism is automatically triggered to ensure that the model always prioritizes the use of the latest and most accurate information.

[0032] Dynamic constraints are applied to several project phases and the corresponding feature data to obtain a material supply time window, including: obtaining the planned start time and preliminary design approval time from the feature data; determining the latest delivery time based on the planned start time, and determining the earliest delivery time based on the preliminary design approval time and a preset limit supply cycle; generating a material supply time window based on the latest delivery time and the earliest delivery time, wherein the material supply time window includes at least a minimum demand value and a maximum demand value.

[0033] This process deeply integrates project schedule constraints with material supply cycles, calculating precise material supply time windows for each project.

[0034] The latest delivery time for the project is calculated by adding two months to the planned start date; the earliest delivery time for the materials is calculated by adding two months to the initial design approval date and the corresponding maximum supply cycle.

[0035] The minimum demand value is the total demand for materials in unfulfilled demand where the latest supply time is equal to the target time node; the maximum demand value is the total demand for materials in unfulfilled demand where the supply window includes the target time node.

[0036] The historical supply data and the material supply time window are input into the constructed demand forecasting model for processing to obtain the material demand forecast value. The construction process is configured to use a two-layer stacked long short-term memory neural network structure, with random deactivation layers added between the layers. The historical supply data and the material supply time window are processed, and the material demand forecast value is output through a fully connected layer.

[0037] Specifically, the demand forecasting model also includes feature attention gates and scale fusion gates.

[0038] The feature data corresponding to the project stages are obtained, and the feature data is input into the feature attention gate to obtain weighted features. The weighted features, the historical supply data, and the material supply time window are input into the two-layer stacked long short-term memory neural network. After processing by the random deactivation layer, the multi-scale hidden layer output is obtained. The multi-scale hidden layer output is input into the scale fusion gate for nonlinear fusion, and the predicted material demand value is output through the fully connected layer.

[0039] The first LSTM layer has 64 neurons, the second LSTM layer has 32 neurons, and a Dropout layer is added between the layers. The inactivation rate is set to 0.2 to alleviate the overfitting problem of the model. Finally, the 12-month material demand prediction results are output through a fully connected layer.

[0040] The model was trained using the Adam adaptive optimizer with an initial learning rate of 0.001. A cosine annealing learning rate scheduling strategy was adopted, with a training batch size of 64 and 200 iterations. The early stopping strategy was set to stop training when the validation set loss did not decrease for 8 consecutive epochs.

[0041] The loss function adopts the weighted multi-scale mean squared error loss function, with a monthly loss weight of 0.3, a quarterly loss weight of 0.4, and an annual loss weight of 0.3. By utilizing the characteristic of long-period errors canceling each other out, the overall prediction accuracy is optimized.

[0042] The loss function can also be the mean squared error (MSE), with the optimization objective being to minimize the deviation between the predicted value and the actual purchase quantity.

[0043] In another embodiment, the demand forecasting model employs an improved LSTM neural network. The standard LSTM network controls the flow of information through input, forget, and output gates, effectively addressing the vanishing gradient problem in long-sequence training. However, in the scenario of power grid material forecasting, the standard LSTM has two significant shortcomings: first, it cannot handle multi-source heterogeneous engineering characteristics; and second, it cannot effectively utilize multi-scale statistical regularities.

[0044] This invention improves upon the standard LSTM cell unit by adding a feature attention gate and a scale fusion gate to the traditional gating structure: the feature attention gate can automatically learn the importance weights of features at different engineering stages and dynamically adjust the contribution of each feature subspace to the prediction results; the scale fusion gate will nonlinearly fuse the prediction results at three scales: monthly, quarterly, and annual, using the stability of long-term predictions to correct the volatility of short-term predictions.

[0045] To verify the effectiveness of the present invention, experiments were conducted using actual procurement data of infrastructure projects of the power grid from January 2021 to December 2024. 110kV and 220kV transformers were used as prediction objects, and the results were compared with existing mainstream prediction methods (standard LSTM, ARIMA, XGBoost).

[0046] Training set: January 2021 to December 2023 Validation set: January 2024 to December 2024 Evaluation metrics: Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE) Table 1 Comparison of prediction accuracy of different methods (110kV transformer) Table 2 Comparison of prediction accuracy of different methods (220kV transformer) Taking 110kV and 220kV transformers as examples, data prediction was performed, and the prediction results are shown below: Table 3. Forecast Results of 110kV Transformers from January to December 2024 Unit: unit Table 4. Forecast Results of 220kV Transformers from January to December 2024 Unit: unit Experimental results show that the method of the present invention is significantly superior to the prior art in all evaluation indicators: 1. Monthly forecast accuracy is improved by approximately 30%-40%, resolving the issue of excessive monthly forecast errors in traditional methods. 2. Quarterly forecast accuracy is improved by approximately 40%-50%, effectively supporting the materials department in formulating quarterly procurement plans. 3. Annual forecast accuracy has been improved by more than 50%, and the annual forecast error has been controlled within 5%, fully meeting the accuracy requirements of the power grid company's annual material planning.

[0047] After obtaining the predicted material demand, the power grid material demand prediction method further includes: acquiring the predicted material demand and the actual delivery data of the target power grid project; determining the prediction deviation rate based on the predicted material demand and the actual delivery data; when the prediction deviation rate exceeds a preset threshold, using the actual delivery data as a new training sample to incrementally learn the demand prediction model, updating the network parameters of the demand prediction model, and obtaining an optimized demand prediction model.

[0048] The predicted demand quantity and predicted delivery time of each category of materials are extracted from the output of the demand forecasting model. At the same time, the actual delivery quantity and actual delivery time of the corresponding category of materials are extracted from the actual warehousing records of the materials management system. The two are then aligned and matched according to the material category and time dimension.

[0049] For each category of materials, calculate the absolute value of the difference between the predicted quantity and the actual quantity delivered, and then divide it by the actual quantity delivered to obtain the relative deviation rate for that category. Then, weight and sum the deviation rates of all categories according to the material weight or the proportion of the amount to obtain the overall prediction deviation rate.

[0050] The actual arrival data of this period, together with the corresponding engineering stage feature data and supply time window, are used to construct a new training sample pair. While keeping the knowledge learned by the original model unchanged, the parameters of the fully connected layer and the randomly deactivated layer between layers of the double-layer stacked LSTM are fine-tuned only with the new samples. The gradient is calculated and the weights are updated through the backpropagation algorithm to obtain the optimized demand prediction model.

[0051] Experimental verification based on actual power grid engineering data shows that the method of the present invention has excellent predictive performance and generalization ability, and can provide reliable decision support for the advance planning and precise allocation of power grid material supply chain.

[0052] Another embodiment of the present invention provides a power grid material demand forecasting system. For details, please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram shown illustrates the structure of a power grid material demand forecasting system according to one embodiment of the present invention. The device includes: Module 11 is used to acquire engineering construction data, material supply plan data, and historical supply data of the target power grid project; The segmentation module 12 is used to segment the target power grid project based on the engineering construction data and the material supply plan data to obtain several project stages and feature data corresponding to the several project stages. Constraint module 13 is used to dynamically constrain and sort out the feature data corresponding to several project stages to obtain the material supply time window; Processing module 14 is used to input the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast value. The construction process is configured to use a two-layer stacked long short-term memory neural network structure, with random deactivation layers added between layers, to process the historical supply data and the material supply time window, and output the material demand forecast value through a fully connected layer.

[0053] Preferably, the target power grid project is divided into several project phases and corresponding feature data, including: Obtain the project time series from the engineering construction data and the planned procurement time nodes from the material supply plan data; Based on the project timeline and the planned procurement timeline, the target power grid project is divided into several project phases, including the pre-feasibility study phase, the feasibility study-preliminary design phase, and the post-preliminary design phase. For each project phase, engineering attribute parameters corresponding to that phase are extracted from the engineering construction data as feature data for that phase; wherein, the engineering attribute parameters include at least the planned investment amount, planned construction period, and material category list for that phase.

[0054] Preferably, the material supply time window is obtained by dynamically constraining and sorting the characteristic data corresponding to the project stages, including: Obtain the planned start time and preliminary design approval time from the feature data; The latest delivery time is determined based on the planned start time, and the earliest delivery time is determined based on the preliminary approval time and the preset maximum supply cycle. Based on the latest arrival time and the earliest arrival time, a material supply time window is generated, which includes at least the minimum demand value and the maximum demand value.

[0055] Preferably, the demand forecasting model further includes a feature attention gate and a scale fusion gate; the step of inputting the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast includes: Obtain feature data corresponding to the aforementioned project stages, and input the feature data into the feature attention gate to obtain weighted features; The weighted features, the historical supply data, and the material supply time window are input into the two-layer stacked long short-term memory neural network. After processing by the random deactivation layer, the multi-scale hidden layer output is obtained. The output of the multi-scale hidden layer is input to the scale fusion gate for nonlinear fusion, and the predicted value of material demand is output through the fully connected layer.

[0056] Preferably, after obtaining the predicted material demand, the power grid material demand forecasting system further includes: Obtain the predicted material demand and the actual delivery data of the target power grid project; Based on the predicted demand for materials and the actual delivery data, the prediction deviation rate is determined; When the prediction deviation rate exceeds a preset threshold, the actual delivery data is used as a new training sample to incrementally learn the demand prediction model, update the network parameters of the demand prediction model, and obtain an optimized demand prediction model.

[0057] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for forecasting power grid material demand, characterized in that, include: Obtain engineering construction data, material supply plan data, and historical supply data for the target power grid project; Based on the engineering construction data and the material supply plan data, the target power grid project is divided into several project stages and characteristic data corresponding to the project stages. Dynamic constraints are applied to several project phases and the corresponding feature data to obtain the material supply time window. The historical supply data and the material supply time window are input into the constructed demand forecasting model for processing to obtain the material demand forecast value. The construction process is configured to use a two-layer stacked long short-term memory neural network structure, with random deactivation layers added between the layers. The historical supply data and the material supply time window are processed, and the material demand forecast value is output through a fully connected layer.

2. The power grid material demand forecasting method as described in claim 1, characterized in that, The process of dividing the target power grid project into several project phases and corresponding feature data for each project phase includes: Obtain the project time series from the engineering construction data and the planned procurement time nodes from the material supply plan data; Based on the project timeline and the planned procurement timeline, the target power grid project is divided into several project phases, including the pre-feasibility study phase, the feasibility study-preliminary design phase, and the post-preliminary design phase. For each project phase, engineering attribute parameters corresponding to that phase are extracted from the engineering construction data as feature data for that phase; wherein, the engineering attribute parameters include at least the planned investment amount, planned construction period, and material category list for that phase.

3. The power grid material demand forecasting method as described in claim 1, characterized in that, The process of dynamically constraining and organizing the characteristic data corresponding to several project stages to obtain the material supply time window includes: Obtain the planned start time and preliminary design approval time from the feature data; The latest delivery time is determined based on the planned start time, and the earliest delivery time is determined based on the preliminary approval time and the preset maximum supply cycle. Based on the latest arrival time and the earliest arrival time, a material supply time window is generated, which includes at least the minimum demand value and the maximum demand value.

4. The power grid material demand forecasting method as described in claim 1, characterized in that, The demand prediction model also includes a feature attention gate and a scale fusion gate; The step of inputting the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast includes: Obtain feature data corresponding to the aforementioned project stages, and input the feature data into the feature attention gate to obtain weighted features; The weighted features, the historical supply data, and the material supply time window are input into the two-layer stacked long short-term memory neural network. After processing by the random deactivation layer, the multi-scale hidden layer output is obtained. The output of the multi-scale hidden layer is input to the scale fusion gate for nonlinear fusion, and the predicted value of material demand is output through the fully connected layer.

5. The power grid material demand forecasting method as described in claim 1, characterized in that, After obtaining the predicted material demand, the power grid material demand prediction method further includes: Obtain the predicted material demand and the actual delivery data of the target power grid project; Based on the predicted demand for materials and the actual delivery data, the prediction deviation rate is determined. When the prediction deviation rate exceeds a preset threshold, the actual delivery data is used as a new training sample to incrementally learn the demand prediction model, update the network parameters of the demand prediction model, and obtain an optimized demand prediction model.

6. A power grid material demand forecasting system, characterized in that, include: The acquisition module is used to acquire engineering construction data, material supply plan data, and historical supply data of the target power grid project; The segmentation module is used to segment the target power grid project based on the engineering construction data and the material supply plan data, to obtain several project stages and feature data corresponding to the project stages. The constraint module is used to dynamically constrain and sort out the characteristic data corresponding to the project stages to obtain the material supply time window. The processing module is used to input the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast value. The construction process is configured to use a two-layer stacked long short-term memory neural network structure, with random deactivation layers added between the layers, to process the historical supply data and the material supply time window, and output the material demand forecast value through a fully connected layer.

7. The power grid material demand forecasting system as described in claim 6, characterized in that, The process of dividing the target power grid project into several project phases and corresponding feature data for each project phase includes: Obtain the project time series from the engineering construction data and the planned procurement time nodes from the material supply plan data; Based on the project timeline and the planned procurement timeline, the target power grid project is divided into several project phases, including the pre-feasibility study phase, the feasibility study-preliminary design phase, and the post-preliminary design phase. For each project phase, engineering attribute parameters corresponding to that phase are extracted from the engineering construction data as feature data for that phase; wherein, the engineering attribute parameters include at least the planned investment amount, planned construction period, and material category list for that phase.

8. The power grid material demand forecasting system as described in claim 6, characterized in that, The process of dynamically constraining and organizing the characteristic data corresponding to several project stages to obtain the material supply time window includes: Obtain the planned start time and preliminary design approval time from the feature data; The latest delivery time is determined based on the planned start time, and the earliest delivery time is determined based on the preliminary approval time and the preset maximum supply cycle. Based on the latest arrival time and the earliest arrival time, a material supply time window is generated, which includes at least the minimum demand value and the maximum demand value.

9. The power grid material demand forecasting system as described in claim 6, characterized in that, The demand prediction model also includes a feature attention gate and a scale fusion gate; The step of inputting the historical supply data and the material supply time window into the constructed demand forecasting model for processing to obtain the material demand forecast includes: Obtain feature data corresponding to the aforementioned project stages, and input the feature data into the feature attention gate to obtain weighted features; The weighted features, the historical supply data, and the material supply time window are input into the two-layer stacked long short-term memory neural network. After processing by the random deactivation layer, the multi-scale hidden layer output is obtained. The output of the multi-scale hidden layer is input to the scale fusion gate for nonlinear fusion, and the predicted value of material demand is output through the fully connected layer.

10. The power grid material demand forecasting system as described in claim 6, characterized in that, After obtaining the predicted material demand, the power grid material demand forecasting system further includes: Obtain the predicted material demand and the actual delivery data of the target power grid project; Based on the predicted demand for materials and the actual delivery data, the prediction deviation rate is determined. When the prediction deviation rate exceeds a preset threshold, the actual delivery data is used as a new training sample to incrementally learn the demand prediction model, update the network parameters of the demand prediction model, and obtain an optimized demand prediction model.

Images