Power grid project cycle simulation method and device based on deep learning

Abstract

The invention discloses a power grid project period simulation method and device based on deep learning. The method comprises the steps of obtaining project attribute data and project environment data of a power grid project; according to the engineering characteristic relationship, performing data marking on the engineering attribute data and the engineering environment data to obtain marked engineering attribute data and marked engineering environment data; in combination with a pre-constructed double-layer network model, performing time sequence analysis on the annotated project attribute data and the annotated project environment data, and determining a project progress index vector corresponding to the project time sequence feature; and according to a preset time granularity, in combination with the project progress index vector, establishing a project progress simulation curve, and completing simulation of a power grid project period. By analyzing the time sequence relationship among various data in the power grid project, a quantitative deviation feedback mechanism of the simulation result and the actual progress is established, and efficient and accurate simulation of the power grid project period is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of engineering simulation, and particularly relates to a power grid engineering cycle simulation method and device based on deep learning. BACKGROUND

[0002] With the continuous expansion of the scale and the improvement of the complexity of power grid engineering, the traditional project management method has problems such as low progress prediction accuracy and resource allocation lag. The power grid engineering cycle involves multiple links such as design, construction, material supply and investment control, and its progress is dynamically affected by multiple factors such as weather, geographical environment and personnel allocation, showing strong time sequence dependence and nonlinear characteristics. The current power grid engineering cycle management relies on static statistical models or empirical formulas, which cannot capture the dynamic change rule of engineering data, resulting in a large deviation between the simulation results and the actual progress.

[0003] The traditional power grid engineering progress simulation method generally uses single data and only uses single indicators such as investment completion rate or construction progress to realize progress prediction, ignoring the correlation analysis of multi-dimensional information such as accounting data and material supply curve, resulting in a lack of comprehensiveness of the simulation model. At the same time, the time sequence feature extraction capability of the traditional simulation method is weak, and the static regression or moving average model cannot effectively capture the long-term dependence and mutation characteristics of the engineering progress, especially when facing sudden events such as extreme weather or supply chain interruption. Moreover, the model has poor adaptability and needs to be manually adjusted to adapt to different voltage level projects, and has insufficient generalization ability and low efficiency.

[0004] Patent application CN120373641A discloses a power engineering construction management method and system based on digital twinning. The method generates a construction process dynamic atlas by analyzing the dependence relationship of the construction process and dynamically injecting, realizes real-time monitoring and adjustment of the construction process. The dynamic atlas is used for environmental adaptability optimization between processes to reduce construction delays caused by environmental changes. At the same time, a multi-modal construction state field is deduced to help identify potential risks, and an abnormal cause and effect backtracking mechanism can trace construction abnormalities, optimize decisions and reduce risks. Through reversible process scheduling optimization and simulation, flexible adjustment in the construction process is realized, and the engineering resilience and response capability are improved.

[0005] The above-mentioned prior art realizes the simulation of the engineering cycle by constructing a dynamic atlas, but the dynamic atlas has limitations in presenting time sequence data and cannot well present the time sequence relationship between data.

[0006] How to ensure the time sequence relationship between various data in the power grid engineering on the basis of considering the dynamic influence of multiple factors on the power grid engineering, establish a quantitative deviation feedback mechanism between the simulation results and the actual progress, and realize dynamic optimization of power grid engineering decisions is a problem to be solved at present. SUMMARY

[0007] In view of the defects in the prior art, the present application provides a power grid engineering cycle simulation method and device based on deep learning, which comprises the following steps:

[0008] In the first aspect, the present application provides a power grid engineering cycle simulation method based on deep learning, which specifically comprises the following steps: Obtaining engineering attribute data and engineering environment data of a power grid engineering; According to the engineering characteristic relationship, data labeling is performed on the engineering attribute data and the engineering environment data to obtain labeled engineering attribute data and labeled engineering environment data; Combining a pre-constructed double-layer network model, time sequence analysis is performed on the labeled engineering attribute data and the labeled engineering environment data, and an engineering progress index vector corresponding to an engineering time sequence feature is determined; According to a preset time granularity, combining the engineering progress index vector, an engineering progress simulation curve is established to complete the simulation of the power grid engineering cycle.

[0009] Further, the engineering attribute data and the engineering environment data are obtained by the following steps: Obtaining original engineering attribute data and original environment data, wherein the original engineering attribute data includes engineering cost data, engineering construction progress data and engineering material usage data, and the original environment data includes engineering meteorological data, engineering geographical data and engineering personnel data; Data cleaning is performed on the original engineering attribute data and the original environment data to obtain first engineering data and first environment data; The first engineering data and the first environment data are format-converted to unify the data format, thereby obtaining the engineering attribute data and the engineering environment data.

[0010] Further, according to the engineering characteristic relationship, data labeling is performed on the engineering attribute data and the engineering environment data to obtain labeled engineering attribute data and labeled engineering environment data, specifically including: According to the relationship between the engineering attribute and / or the engineering environment in the data sample set, the engineering characteristic relationship is determined; According to the engineering characteristic relationship, a derivable label function is generated; The engineering attribute data and the engineering environment data are marked with the derivable label function to obtain marked engineering attribute data and marked engineering environment data.

[0011] Further, the derivable label function is generated according to the engineering feature relationship, and specifically includes the following steps: According to the feature type, the relationship type and the action parameter in the engineering feature relationship, a label calculation logic expression is constructed, wherein the action parameter includes an influence coefficient and / or a constraint threshold; Based on the label calculation logic expression, a preset function template library is matched to give the derivable label function, wherein the feature type is mapped to a function input variable in the derivable label function, and the action parameter is instantiated as a function calculation parameter.

[0012] Further, the marked engineering attribute data and the marked engineering environment data are subjected to time sequence analysis in combination with a pre-constructed double-layer network model, and an engineering time sequence feature corresponding engineering progress index vector is determined, specifically including: The marked engineering attribute data and the marked engineering environment data are subjected to normalization processing to obtain standard engineering attribute information and standard engineering environment information; The standard engineering attribute information and the standard engineering environment information are input into a causal perception LSTM layer for time sequence analysis, time sequence features are extracted, and a time sequence feature vector is obtained; The time sequence feature vector is mapped in a full connection layer to obtain an engineering progress index vector.

[0013] Further, the double-layer network model includes a causal perception LSTM layer and a full connection layer, and the causal perception LSTM layer and the full connection layer are connected through a dropout unit, wherein the causal perception LSTM layer includes an LSTM module and a causal perception module. The construction of the double-layer network model specifically includes: The historical attribute data and the historical environment data of the power grid project are obtained; The historical attribute data and the historical environment data of the power grid project are subjected to normalization processing to obtain standard historical engineering data; The initial causal perception LSTM layer is iteratively trained according to the standard historical engineering data, and the model parameters of the initial causal perception LSTM layer are adjusted to obtain an intermediate-state causal perception LSTM layer in combination with the current iteration number; The intermediate-state causal perception LSTM layer and the full connection layer are trained in combination with a loss function to obtain a primary double-layer network model; The primary double-layer network model is verified according to historical engineering verification data to obtain an engineering progress deviation rate of the primary double-layer network model; Based on the engineering progress deviation rate, the discarding parameters of the discarding unit are adjusted, the primary double-layer network model is re-verified, and the double-layer network model is obtained.

[0014] Further, the initial causal perception LSTM layer is iteratively trained according to standard historical engineering data, the model parameters of the initial causal perception LSTM layer are adjusted in combination with the current iteration number to obtain an intermediate-state causal perception LSTM layer, and the specific operations include the following steps. According to the current iteration number, a current iteration coefficient of the current iteration number and the maximum iteration number is calculated. By judging the current iteration coefficient, the causal structure parameters of the initial causal perception LSTM layer are frozen. According to the standard historical engineering data, the initial causal perception LSTM layer is trained, the time sequence dynamic parameters in the initial causal perception LSTM layer are optimized, and the intermediate-state causal perception LSTM layer is obtained.

[0015] Further, the time sequence dynamic parameters in the initial causal perception LSTM layer are optimized to obtain the intermediate-state causal perception LSTM layer, and the specific operations include the following steps. Based on the causal perception module, counterfactual time sequence is generated, the counterfactual time sequence is mixed with historical engineering data to form a training batch, the weighted sum of the prediction loss and the causal consistency loss is used to update the time sequence dynamic parameters. Until the convergence condition is met, the causal structure parameters and the time sequence dynamic parameters are solidified, and the intermediate-state causal perception LSTM layer is obtained.

[0016] Further, the primary double-layer network model is verified according to historical engineering verification data, and the engineering progress deviation rate of the primary double-layer network model is obtained, and the specific operations include the following steps. The verification data in the historical engineering verification data is input into the primary double-layer network model to obtain historical engineering progress prediction data. The difference between the historical engineering progress actual data and the historical engineering progress prediction data in the historical engineering verification data is calculated to obtain an engineering progress verification difference. According to the ratio of the engineering progress verification difference to the historical engineering progress actual data, all the engineering progress verification differences are fused to obtain the engineering progress deviation rate of the primary double-layer network model.

[0017] In the second aspect, the application further provides a power grid engineering cycle simulation device based on deep learning, which adopts the power grid engineering cycle simulation method based on deep learning in any one of the above aspects, and includes the following. A data acquisition module is configured to acquire engineering attribute data and engineering environment data of a power grid project. A data labeling module is configured to label the engineering attribute data and the engineering environment data according to an engineering feature relationship to obtain labeled engineering attribute data and labeled engineering environment data. The progress analysis module is configured to perform time sequence analysis on the labeled engineering attribute data and the labeled engineering environment data in combination with the pre-constructed double-layer network model, and determine an engineering progress index vector corresponding to an engineering time sequence feature; The progress simulation module is configured to establish an engineering progress simulation curve in combination with the engineering progress index vector according to a preset time granularity, and complete simulation of the power grid engineering cycle.

[0018] The power grid engineering cycle simulation method and device based on deep learning provided by the application have at least the following beneficial effects: (1) By constructing a double-layer network model, multi-dimensional time sequence data of a power grid engineering is integrated, time sequence dependence is captured by using a gating mechanism, causal perception is strengthened, feature correlation is associated, a bias feedback optimization model is combined, and the reliability and adaptability of the full-cycle simulation of the power grid engineering are improved.

[0019] (2) The causal perception module formalizes the prior causal relationship in the power grid engineering and injects it into the model training process, so that the prediction result of the model is traceable; the LSTM module captures the time correlation in the power grid engineering data and adapts to the dynamic changes of the power grid engineering. The combination of the two improves the prediction accuracy and generalization ability of the model.

[0020] (3) By using counterfactual data and causal consistency loss, the model is forced to follow the prior causal logic of the engineering, avoiding the situation that the model only fits the data but does not follow the logic; by using prediction loss and back propagation, time sequence dynamic parameters are optimized, so that the model adapts to the time sequence features of the historical data; the causal structure parameters are fixed, only the time sequence dynamic parameters are updated, and the balance between prior knowledge and data driving is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The flowchart of the power grid engineering cycle simulation method based on deep learning provided by the embodiment of the application is shown in the figure; Figure 2 The flowchart of the data labeling provided by the embodiment of the application is shown in the figure; Figure 3 The flowchart of determining the engineering progress index vector provided by the embodiment of the application is shown in the figure; Figure 4 The model architecture diagram of the double-layer network model provided by the embodiment of the application is shown in the figure; Figure 5 The structural block diagram of the power grid engineering cycle simulation device based on deep learning provided by the embodiment of the application is shown in the figure.

[0022] Among them, 201, data acquisition module; 202, data labeling module; 203, progress analysis module; 204, progress simulation module. DETAILED DESCRIPTION

[0023] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings of the specification and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor belong to the scope of protection of the present application.

[0024] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise. "Multiple" generally includes at least two.

[0025] It should also be noted that the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the goods or devices including a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such goods or devices. Without more limitation, the element defined by the sentence "including a…" does not exclude the presence of another identical element in the goods or devices including the element.

[0026] In the process of power system construction, there are many construction links, long construction period, complex and changeable construction environment, and high coupling relationship between multi-dimensional indicators such as engineering quality, safety and progress. At present, the traditional power engineering construction management mode generally relies on static two-dimensional drawings, manual on-site supervision, offline communication and coordination and other means, and lacks dynamic modeling and real-time sensing ability for the whole life cycle of the project. On the one hand, the traditional method is difficult to realize the data connection of design, construction and operation and maintenance stage, leading to serious information fragmentation and data island problem, and unable to grasp the actual progress and construction quality status of the project in time; on the other hand, the project management decision-making process is seriously dependent on the experience of management personnel, lacks scientific support based on multi-source data fusion analysis, and is easy to appear problems such as project node delay, unreasonable construction resource allocation, late discovery of quality hidden danger, etc., which directly affects the construction efficiency and operation reliability of the project.

[0027] In addition, in the face of complex and changeable construction site environment, the traditional method lacks effective technical means in engineering simulation, process optimization and construction simulation, and is difficult to respond quickly and dynamically adjust to the sudden situation in the engineering process, reducing the risk response ability and system resilience of the project. Especially in large-scale engineering projects across regions and across professions, there are many participating units, scattered data sources, and complex management system, so it is more difficult for the traditional method to realize the collaborative perception and intelligent management and control of the whole engineering system.

[0028] The rise of deep learning technology has provided new ideas for processing complex time series data. Among them, LSTM has shown advantages in power load forecasting and equipment fault diagnosis due to its gating mechanism, but it has not yet been systematically applied to the field of full-cycle progress simulation of power grid engineering.

[0029] Based on this, the present invention provides a deep learning-based method and apparatus for simulating the cycle of power grid engineering projects. The method includes: acquiring engineering attribute data and engineering environment data of power grid engineering projects; labeling the engineering attribute data and engineering environment data according to engineering feature relationships to obtain labeled engineering attribute data and labeled engineering environment data; performing time series analysis on the labeled engineering attribute data and labeled engineering environment data using a pre-constructed two-layer network model, and determining the engineering progress index vector corresponding to the engineering time series characteristics; and establishing an engineering progress simulation curve based on a preset time granularity and the engineering progress index vector to complete the simulation of the power grid engineering cycle.

[0030] like Figure 1 As shown in the figure, this embodiment of the invention provides a deep learning-based method for simulating the cycle of power grid engineering, with the following specific steps: S101: Obtain engineering attribute data and engineering environment data for power grid projects.

[0031] Furthermore, the engineering attribute data and engineering environment data are obtained through the following steps: Obtain raw project attribute data and raw environmental data. The raw project attribute data includes project cost data, project construction progress data, and project material usage data. The raw environmental data includes project meteorological data, project geographical data, and project personnel data. Data cleaning is performed on the original engineering attribute data and original environmental data to obtain the first engineering data and the first environmental data. The first engineering data and the first environmental data are converted to a unified format to obtain engineering attribute data and engineering environmental data.

[0032] In a specific example, the target project is a power distribution network upgrade and renovation project with a project cycle of 6 months. It requires the replacement of 20 distribution towers, the upgrading of 5 power lines, and the equipment renewal of 3 substations. The core control objective is to ensure that the schedule deviation does not exceed 5%. A deep learning-based full-cycle simulation method for power grid engineering is used to achieve accurate prediction and control of the project schedule. First, four types of core data for the target project are collected: raw project attribute data including project cost data, project construction progress data, and project material usage data. Project cost data covers the records of funds received at each stage, such as equipment procurement, construction service fees, design and supervision fees. Project construction progress data includes percentage data on the progress of tower installation, line laying, and substation equipment renewal. Project material usage data involves the supply amount, delivery time, and inventory balance of key materials such as transformers, cables, and towers. Also included is raw environmental data associated with the raw project attribute data, including meteorological data, geographical data, and personnel data. Specifically, this includes meteorological data during construction, such as the number of days with rainfall, high temperatures, and strong winds; geographical data of the construction area, such as road distribution and underground pipeline locations; and personnel configuration data, including on-duty time and skill levels.

[0033] All original project attribute data and original environment data are cleaned, including removing duplicate data and filling in missing values, to obtain the first project data and the first environment data. Then, the data format is converted to unify the data format, resulting in project attribute data and project environment data.

[0034] In a specific example, all collected initial project data and initial environmental data are stored uniformly in CSV format. This format includes timestamp data type labels and numerical units. The timestamp uses days as the time granularity to accurately correspond to daily construction progress. The data type labels respectively indicate the accounting data, investment completion data, actual construction progress data, material curve data, meteorological data, geographical environment data, and construction personnel configuration data, which facilitates subsequent model recognition. The numerical units are matched according to the data type: the numerical unit for funds is ten thousand yuan, the numerical unit for progress is %, the numerical unit for meteorology is days, and the numerical unit for personnel is people.

[0035] S102: Based on the engineering characteristic relationships, label the engineering attribute data and engineering environment data to obtain labeled engineering attribute data and labeled engineering environment data.

[0036] Furthermore, referring to Figure 2 Based on the relationships between engineering characteristics, engineering attribute data and engineering environmental data are labeled to obtain labeled engineering attribute data and labeled engineering environmental data, specifically including: Based on the relationship between engineering attributes and / or engineering environment in the data sample set, determine the engineering characteristic relationships; According to the engineering characteristic relationship, a derivable label function is generated; In combination with the derivable label function, data labeling is performed on the engineering attribute data and the engineering environment data to obtain labeled engineering attribute data and labeled engineering environment data.

[0037] It can be understood that the engineering characteristic relationship is the association relationship between the engineering attribute data and the engineering environment data, which is structuredly represented by a knowledge graph: the nodes of the knowledge graph represent characteristic types, including engineering attribute characteristics and engineering environment characteristics; the edges of the knowledge graph represent relationship types, including causal relationships, constraint relationships, association relationships, etc.; and the attributes of the edges represent action parameters, i.e., quantitative indicators of relationships, including constraint thresholds (such as "3 days" in "rainy days > 3 days") and influence coefficients (such as "10%" in "construction progress decreases by 10%").

[0038] In a specific example, the engineering attribute data and the engineering environment data in the above-described CSV format are subjected to data labeling, and the complete data of the target project in the previous 2 months are selected as the labeling data. Based on the relationship between the engineering attribute and / or the engineering environment in the labeling data, the engineering characteristic relationship is obtained, for example, the construction progress is improved by 5%-8% within 3 days after the material supply is in place, and the investment completion rate is decreased by 10% in the month when the rainy days exceed 3 days.

[0039] Based on the engineering characteristic relationship of "material in place → progress jump within 3 days", the relationship is specifically represented as:

[0040] Wherein, Mat(t)=1 is the value v mat (t)≥θ mat , P(t) is the value of the "actual construction progress" label on the t-th day (unit: %).

[0041] Based on the engineering characteristic relationship of "continuous rainfall → monthly investment decrease", the relationship is specifically represented as:

[0042] Wherein, Rain(m) is the number of rainy days in the m-th month, I finish (m) is the total amount of completed investment in the m-th month, I plan (m) is the total amount of planned investment in the m-th month.

[0043] Further, according to the engineering characteristic relationship, a derivable label function is generated, specifically including: According to the characteristic type, the relationship type and the action parameter in the engineering characteristic relationship, a label calculation logic expression is constructed, wherein the action parameter includes an influence coefficient and / or a constraint threshold. The label calculation logic expression is matched with a preset function template library to obtain a derivable label function, wherein the derivable label function maps the feature type to a function input variable and instantiates the action parameter as a function calculation parameter.

[0044] It should be noted that the derivable label function is to convert the engineering feature relationship into a calculable mathematical rule for automatic labeling data. The label calculation logic expression is a rule for describing the feature relationship in a formal language, which needs to include the feature type, the relationship type and the action parameter. For a causal relationship (such as “rainfall leads to progress decline”), a “condition-adjustment” structure is adopted; for a constraint relationship (such as “investment amount must be greater than or equal to material procurement cost”), a “constraint-labeling” structure is adopted; and for a multi-feature association (such as “rainfall + insufficient investment leads to progress decline”), a “multi-condition-joint adjustment” structure is adopted. The function template library is a predefined standardized function structure for converting the logic expression into an executable formula, and each template includes an input variable placeholder, a calculation parameter placeholder and a logic structure. The feature type (such as “rainfall days”) in the logic expression is mapped to the input variable in the template, and the action parameter (such as “3 days” and “15%”) in the logic expression is filled into the calculation parameter placeholder of the template, and through replacing the placeholders in the template, a specific function, i.e., the derivable label function, is obtained. By inputting any unlabeled engineering attribute data or engineering environment data (t′, L′, u′) into the derivable label function, labeled engineering attribute data and labeled engineering environment data can be obtained.

[0045] In a certain example, the derivable label function is specifically represented as:

[0046] wherein m(t′) is the month to which the t′th day belongs, t′ is the construction time progress of the target project, which is used to accurately locate the day on which the engineering attribute data or the engineering environment data occurs in the entire construction period, L′ is the numerical value of the engineering attribute data or the engineering environment data, such as accounting data, material curve and meteorological data, and u′ is the unit corresponding to L′.

[0047] S103: In combination with the pre-constructed double-layer network model, time sequence analysis is performed on the labeled engineering attribute data and the labeled engineering environment data, and an engineering progress index vector corresponding to the engineering time sequence feature is determined.

[0048] Further, in combination with the pre-constructed double-layer network model, time sequence analysis is performed on the labeled engineering attribute data and the labeled engineering environment data, and an engineering progress index vector corresponding to the engineering time sequence feature is determined, referring to Figure 3 , and specifically comprising: The labeled engineering attribute data and the labeled engineering environment data are normalized to obtain standard engineering attribute information and standard engineering environment information; The standard engineering attribute information and the standard engineering environment information are input into a causal perception LSTM layer for time sequence analysis to extract time sequence features and obtain a time sequence feature vector; The time sequence feature vector is mapped in a full connection layer to obtain an engineering progress index vector.

[0049] In a specific embodiment, a double-layer network model suitable for a target power grid project is constructed, the double-layer network model comprising a long short-term memory network model, and a full connection layer is added to a top layer of the long short-term memory network model to form a double-layer architecture of LSTM feature extraction plus full connection layer index conversion. Normalization can be performed in a Min-Max normalization manner or a Z-score normalization manner. Through normalization processing of labeled data, dimensional differences are eliminated, and training imbalance caused by value range differences of the model is avoided.

[0050] The causal perception LSTM layer comprises an LSTM module and a causal perception module. The causal perception module combines engineering feature relationships in a knowledge graph to perform weighted enhancement on an input vector of the LSTM module. The LSTM module extracts time sequence features through a synergistic effect of an input gate, a forgetting gate, a memory cell and an output gate, and is specifically expressed as:

[0051] wherein, i t is a state value of the input gate at t moment, used to control an update proportion of the current standard engineering data to the memory cell, σ is a sigmoid activation function, w i is a weight matrix of the input gate acting on the current standard engineering data, x t is the standard engineering data at t moment, containing an account amount, an investment completion rate, a construction progress percentage and a material supply amount, U i is a weight matrix of the input gate acting on the model output at t-1 moment, h t-1 is an output vector of the long short-term memory network layer at t-1 moment, reflecting historical time sequence features, b i is a bias vector of the input gate; f t is a state value of the forgetting gate at t moment, used to control a reservation proportion of historical memory information, w f is a weight matrix of the forgetting gate acting on the current standard engineering data, U f is a weight matrix of the forgetting gate acting on the model output at t-1 moment, b f is a bias vector of the forgetting gate; C t is a state value of the memory cell at t moment, used to store fused historical and current information, and is a vector element level multiplication operator, C t-1is the state value of the memory cell at time t-1, tanh() is the hyperbolic tangent activation function, w c is the weight matrix of the memory cell acting on the current standard engineering data, U c is the weight matrix of the memory cell acting on the model output at time t-1, b c is the bias vector of the memory cell; o t is the state value of the output gate at time t, used to control the transmission ratio of memory cell information to the current output, w o is the weight matrix of the output gate acting on the current standard engineering data, U o is the weight matrix of the output gate acting on the model output at time t-1, b o is the bias vector of the output gate.

[0052] The input gate controls the update ratio of the current data to the memory cell, ensuring that the latest key data such as the amount of daily material arrival and the number of construction personnel can be integrated into the model in a timely manner. The forget gate is used to filter historical memory information, eliminating invalid historical data such as temporary adjustments of construction personnel that do not affect progress, and retaining effective information such as stable progress after the full amount of monthly investment is received. The memory cell stores the fused historical and current information, forming a time series correlation throughout the entire project cycle. The output gate controls the transmission ratio of memory cell information to the current output, preferentially outputting features strongly related to progress such as time correlation features of material supply and tower installation. Finally, the causal perception LSTM layer outputs a time series feature vector containing time-dependent information of the project progress.

[0053] After the causal perception LSTM layer extracts the time series features and outputs, the extracted features are converted into the project progress indicators of the target project through the top fully connected layer and the feature indicator conversion formula, which specifically includes the actual construction completion rate, the investment and construction progress difference value, and the material supply matching degree, providing intuitive indicators for subsequent progress control.

[0054] The conversion relationship between the time series feature vector and the project progress indicators is specifically represented as:

[0055] wherein, is the project progress indicator vector output by the fully connected layer at time t, including the construction completion rate, the investment and construction progress difference value, and the material supply matching degree, w is the weight matrix of the fully connected layer, h t is the time series feature vector output by the causal perception LSTM layer at time t-1, b is the bias vector of the fully connected layer, and the dimension of the bias vector is the same as the number of project progress indicators.

[0056] Further, with reference to Figure 4The double-layer network model comprises a causal perception LSTM layer and a full connection layer, and the causal perception LSTM layer and the full connection layer are connected through a dropout unit, wherein the causal perception LSTM layer comprises an LSTM module and a causal perception module; The double-layer network model is constructed, and specifically includes the following steps: The historical attribute data and the historical environment data of the power grid project are obtained; The historical attribute data and the historical environment data of the power grid project are normalized to obtain standard historical engineering data; The initial causal perception LSTM layer is iteratively trained according to the standard historical engineering data, and the model parameters of the initial causal perception LSTM layer are adjusted to obtain an intermediate causal perception LSTM layer in combination with the current iteration number; The intermediate causal perception LSTM layer and the full connection layer are trained in combination with a loss function to obtain a primary double-layer network model; The primary double-layer network model is verified according to historical engineering verification data to obtain an engineering progress deviation rate of the primary double-layer network model; Based on the engineering progress deviation rate, the dropout parameters of the dropout unit are adjusted, the primary double-layer network model is reverified, and a double-layer network model is obtained.

[0057] In a specific example, historical engineering data for training the model is collected and normalized to obtain standard historical engineering data, covering full-cycle information of engineering attributes and environment. The parameters of the initial causal perception LSTM layer, including the input gate weight matrix, the forget gate weight matrix, the memory cell weight matrix, and the output gate weight matrix, are randomly initialized. The standard historical engineering data is then used to train the initial causal perception LSTM layer: the parameters of the first two layers of the LSTM are frozen for the first 10 iterations, and the top-layer parameters are focused on training; 50% of the bottom-layer parameters are unfrozen for the 10th to 20th iterations, allowing partial bottom-layer parameters to be fine-tuned according to historical data; and after 20 iterations, all parameters are completely unfrozen to allow the model to adapt to the characteristics of the historical data. The mean square error (MSE) is used as the loss function to minimize the error between the LSTM output and the historical time-series features, and the intermediate-state causal perception LSTM layer is obtained. The intermediate-state causal perception LSTM layer is connected to the fully connected layer, the mean square error and L2 regularization are used as the joint loss function, and the Adam optimizer is used to update the model parameters through backpropagation, obtaining the primary double-layer network model, wherein the model parameters include the weight matrix of the causal perception LSTM layer, the weight matrix of the fully connected layer, and the bias vector. To evaluate the prediction accuracy of the primary double-layer network model, the standard historical engineering data not involved in training is selected as the validation set, input into the primary model, and the engineering progress deviation rate is calculated. The model parameters are adjusted according to the deviation rate: if the deviation rate is too high, the dropout probability of the dropout unit is reduced to enhance the model's memory ability; if the deviation rate is too low but overfitting, the dropout probability of the dropout unit is increased to enhance the model's generalization ability. The primary model is retrained with the adjusted dropout probability, and the deviation rate is verified again to obtain the final double-layer network model, which has a deviation rate that meets the specified threshold and can be used for time-series analysis.

[0058] Further, the initial causal perception LSTM layer is iteratively trained according to the standard historical engineering data, and the model parameters of the initial causal perception LSTM layer are adjusted to obtain the intermediate-state causal perception LSTM layer according to the current iteration number, specifically including: According to the current iteration number, the current iteration coefficient of the current iteration number and the maximum iteration number is calculated; By judging the current iteration coefficient, the causal structure parameters of the initial causal perception LSTM layer are frozen; According to the standard historical engineering data, the initial causal perception LSTM layer is trained to optimize the time-series dynamic parameters in the initial causal perception LSTM layer, and the intermediate-state causal perception LSTM layer is obtained.

[0059] In a specific embodiment, the model parameters of the double-layer network model are adjusted according to the data characteristics of the target project. Specifically, according to the amount and complexity of the project data, the maximum number of iterations is set, the ratio of the current number of iterations to the maximum number of iterations is calculated as the current iteration coefficient. According to the current iteration coefficient, the freezing proportion of the core parameters such as causal weight, constraint threshold and causal correlation matrix in the causal perception module is determined. With the increase of the iteration coefficient, the freezing proportion linearly decreases. Then through the training of the standard historical project data, the weight matrices in the LSTM module are updated through back propagation to obtain the intermediate causal perception LSTM layer.

[0060] The model loss function adopts a combined form including a mean square error term and an L2 regularization term. The L2 regularization loss function balances the model fitting degree and the generalization ability, avoiding the model from excessively relying on the abnormal material supply data of a few days. The model loss function is specifically represented as:

[0061] wherein L is the total loss value of the model, is the mean square error term, n is the number of training samples in a single batch, y t is the actual value of the progress index at time t; is the output of the full connection layer at time t, i.e. the predicted value of the progress index at time t, λ is a regularization coefficient, θ k is the trainable parameter of the model. In a specific example, to prevent model overfitting, λ is set to 0.0005, λ ∈ (0001, 0.01).

[0062] In a specific example, a dropout unit is set between the long short-term memory network layer and the full connection layer, which is used to discard part of the output results of the long short-term memory network layer. For example, the dropout parameter of the dropout unit is set to 0.2, which means that 20% of the neurons in the long short-term memory network layer are randomly masked, avoiding the model from excessively relying on the features extracted by specific neurons, and improving the adaptability to the fluctuations of project data such as progress fluctuations caused by temporary rainfall.

[0063] A batch normalization layer is also set before the long short-term memory network layer, which adopts a standardization formula to standardize each batch of project data, eliminating the influence of data dimension difference on model training.

[0064] Further, the time sequence dynamic parameters in the initial causal perception LSTM layer are optimized to obtain the intermediate causal perception LSTM layer, which includes the following steps: Based on the causal perception module, counterfactual time sequence is generated, the counterfactual time sequence is mixed with historical project data to form a training batch, and the weighted sum of prediction loss and causal consistency loss is used to update the time sequence dynamic parameters; Until the convergence condition is met, the solidified causal structure parameters and the timing dynamic parameters are obtained, and the intermediate causal perception LSTM layer is obtained.

[0065] It can be understood that the causal perception module has stored engineering prior causal relationships based on the knowledge graph, and the counterfactual time sequence refers to modifying the causal variables (such as "rainy days") in the historical data, and generating "hypothetical" time series data (such as "if the rainy days are ≤3 days, what will the progress be") according to the prior relationship of the causal perception module. First, the causal driving variables (such as "rainy days", "material supply amount", etc.) are extracted from the standard historical engineering data; then, based on the prior relationship of the causal perception module, a counterfactual scenario (such as "if the rainy days are reduced from 5 days to 2 days, what will the progress be") is defined; according to the influence coefficient of the causal relationship (such as the influence coefficient of "rainy days > 3 days → progress decreases by 10%" is 0.1), the result variable under the counterfactual scenario is calculated, and the result variable under different time sequences forms the counterfactual time sequence.

[0066] The counterfactual data and the historical data are mixed in a ratio of 1:4 to 1:5, and the mixed data set is divided into batches in chronological order to maintain the time sequence continuity. The prediction loss is the error between the result predicted by the model for the historical data and the true result, which measures the fitting ability of the model to the true data; the causal consistency loss is the error between the result predicted by the model for the counterfactual data and the counterfactual label generated by the causal perception module, which measures the degree to which the model follows the causal logic. The importance of the two losses is adjusted by weight, and the sum is taken as the weighted total loss. The Adam optimizer is used to optimize the weight matrix and bias vector of the LSTM module through back propagation, and the prior parameters (influence coefficient, constraint threshold, etc.) of the causal perception module are not involved in the back propagation. When the weighted total loss decreases by less than 10 -4 or reaches the preset maximum number of iterations, the model converges, the training stops, the parameters are solidified, and the intermediate causal perception LSTM layer is obtained, which can be processed and meets the causal logic.

[0067] Through the counterfactual data and the causal consistency loss, the model is forced to follow the engineering prior causal logic, avoiding the model only fitting the data without following the logic; through the prediction loss and the back propagation, the timing dynamic parameters are optimized, allowing the model to adapt to the timing characteristics of the historical data; the causal structure parameters are solidified, only the timing dynamic parameters are updated, achieving a balance between prior knowledge and data-driven, and improving the accuracy and efficiency of power grid engineering cycle simulation.

[0068] Further, according to the historical engineering verification data, the primary double-layer network model is verified to obtain the engineering progress deviation rate of the primary double-layer network model, specifically including: The verification data in the historical engineering verification data is input into the primary double-layer network model to obtain historical engineering progress prediction data; The difference between the actual data of the historical engineering progress and the predicted data of the historical engineering progress in the historical engineering verification data is calculated to obtain an engineering progress verification difference value. According to the ratio of the engineering progress verification difference value to the actual data of the historical engineering progress, all engineering progress verification difference values are fused to obtain an engineering progress deviation rate of the primary double-layer network model.

[0069] In a specific embodiment, the standard historical engineering data is divided into a training set and a verification set, and in this example, the ratio of the training set to the test set is 7:3. The verification set is input into the primary double-layer network model, the test set covers complete data of the target engineering for 2 months after implementation, and the primary double-layer network model outputs a full-cycle progress curve of the target engineering, which includes predicted values of three indexes of the actual construction completion rate, the construction progress difference value and the material supply matching degree. The deviation rate is calculated to compare the historical engineering progress prediction data output by the primary double-layer network model with the actual data of the historical engineering progress, and the deviation rate of the two is calculated, which is specifically represented as:

[0070] Wherein, the deviation rate is the average relative error of the historical engineering progress prediction data and the actual data of the historical engineering progress, T is the time step of the target engineering full cycle, is the historical engineering progress prediction data at time t; is the actual data of the historical engineering progress at time t.

[0071] In the specified requirements of this example, the predetermined threshold of the 110kV power grid engineering is ≤5%, and the predetermined threshold of the 220kV and above power grid engineering is ≤3%. During the first verification, the model fails to capture the feature that the line laying is suspended for 3 days due to rain, and the deviation rate is 7.2%, which exceeds the predetermined threshold of the 110kV power grid engineering.

[0072] In view of the problem of excessive deviation, the model parameters of the primary double-layer network model and the dropout parameters of the dropout unit are readjusted, for example, the dropout parameter is lowered from 0.2 to 0.15 to enhance the model's feature memory of short-term abnormal data; at the same time, the L2 regularization coefficient is adjusted to 0.001 to improve the model's fault tolerance to abnormal data. After retraining the model, the deviation rate is reduced to 4.1%, which meets the predetermined threshold requirement, and the model can be used for subsequent progress simulation and control of the target engineering.

[0073] S104: According to the preset time granularity, an engineering progress simulation curve is established combining the engineering progress index vector, and the simulation of the power grid engineering cycle is completed.

[0074] After the double-layer network model is trained, the double-layer network model is used to analyze the real-time input engineering attribute data and engineering environment data, generate corresponding engineering progress index vectors, generate engineering progress simulation curves according to time granularity, and complete simulation of the power grid engineering cycle. For example, if the time granularity is a day, the corresponding engineering progress index vector is output according to the engineering attribute data and the engineering environment data of each day, and the engineering progress index vectors of different stages form the engineering progress simulation curve.

[0075] With reference to Figure 5 The embodiment of the present application provides a power grid engineering cycle simulation device based on deep learning, which comprises: A data acquisition module 201 is configured to acquire engineering attribute data and engineering environment data of a power grid engineering. A data labeling module 202 is configured to label the engineering attribute data and the engineering environment data according to engineering feature relationships to obtain labeled engineering attribute data and labeled engineering environment data. A progress analysis module 203 is configured to combine a pre-constructed double-layer network model to perform time sequence analysis on the labeled engineering attribute data and the labeled engineering environment data, and determine an engineering progress index vector corresponding to engineering time sequence features. A progress simulation module 204 is configured to combine the engineering progress index vector to establish an engineering progress simulation curve according to a preset time granularity, and complete simulation of the power grid engineering cycle.

[0076] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described modules can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0077] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to these embodiments once they understand the basic creative concept. Therefore, the appended claims are intended to include the preferred embodiments and all changes and modifications falling within the scope of the present application. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.

Claims

1. A deep learning-based method for cyclic simulation of power grid engineering, characterized in that, include: Obtain engineering attribute data and engineering environment data for power grid projects; Based on the engineering characteristic relationships, the engineering attribute data and engineering environment data are labeled to obtain labeled engineering attribute data and labeled engineering environment data; By combining a pre-built two-layer network model, time series analysis is performed on labeled engineering attribute data and labeled engineering environment data, and the engineering progress index vector corresponding to the engineering time series characteristics is determined. Based on the preset time granularity and combined with the project progress index vector, a project progress simulation curve is established to complete the simulation of the power grid project cycle.

2. The deep learning-based power grid engineering cycle simulation method as described in claim 1, characterized in that, Project attribute data and project environment data are obtained through the following steps: Obtain raw project attribute data and raw environmental data. The raw project attribute data includes project cost data, project construction progress data, and project material usage data. The raw environmental data includes project meteorological data, project geographical data, and project personnel data. Data cleaning is performed on the original engineering attribute data and original environmental data to obtain the first engineering data and the first environmental data. The first engineering data and the first environmental data are converted to a unified format to obtain engineering attribute data and engineering environmental data.

3. The deep learning-based power grid engineering cycle simulation method as described in claim 1, characterized in that, Based on the relationships between engineering characteristics, engineering attribute data and engineering environment data are labeled to obtain labeled engineering attribute data and labeled engineering environment data, specifically including: Based on the relationship between engineering attributes and / or engineering environment in the data sample set, determine the engineering characteristic relationships; Generate a derivable label function based on engineering feature relationships; By combining a derivable label function, engineering attribute data and engineering environment data are labeled to obtain labeled engineering attribute data and labeled engineering environment data.

4. The deep learning-based power grid engineering cycle simulation method as described in claim 3, characterized in that, Based on engineering feature relationships, a derivable label function is generated, which includes the following steps: Based on the feature type, relationship type, and action parameters in the engineering feature relationship, construct the label calculation logic expression, where the action parameters include the influence coefficient and / or the constraint threshold; Based on the label calculation logical expression, a pre-set function template library is matched to give a deducible label function. In the deducible label function, the feature type is mapped to the function input variable and the applied parameter is instantiated as the function calculation parameter.

5. The deep learning-based power grid engineering cycle simulation method as described in claim 1, characterized in that, By combining a pre-constructed two-layer network model, time-series analysis is performed on labeled engineering attribute data and labeled engineering environment data, and the engineering progress index vector corresponding to the engineering time-series characteristics is determined, specifically including: Normalize the labeled project attribute data and labeled project environment data to obtain standard project attribute information and standard project environment information; Standard engineering attribute information and standard engineering environmental information are input into the causal perception LSTM layer for time series analysis, and time series features are extracted to obtain time series feature vectors; The time-series feature vector is mapped in the fully connected layer to obtain the project progress index vector.

6. The deep learning-based power grid engineering cycle simulation method as described in claim 5, characterized in that, The two-layer network model includes a causal sensing LSTM layer and a fully connected layer, which are connected by dropout units. The causal sensing LSTM layer includes an LSTM module and a causal sensing module. The construction of a two-layer network model specifically includes: Acquire historical attribute data and historical environmental data of power grid projects; The historical attribute data and historical environmental data of power grid projects are normalized to obtain standard historical project data. The initial causal perception LSTM layer is iteratively trained based on standard historical engineering data. The model parameters of the initial causal perception LSTM layer are then adjusted based on the current iteration number to obtain the intermediate causal perception LSTM layer. By combining the loss function, the intermediate causal awareness LSTM layer and the fully connected layer are trained to obtain the primary two-layer network model; Based on historical engineering verification data, the primary two-layer network model was verified, and the project schedule deviation rate of the primary two-layer network model was obtained. Based on the project schedule deviation rate, the discard parameters of the discarded units are adjusted, and the primary two-layer network model is re-verified to obtain the two-layer network model.

7. The deep learning-based power grid engineering cycle simulation method as described in claim 6, characterized in that, The initial causal sensing LSTM layer is iteratively trained based on standard historical engineering data. Then, by adjusting the model parameters of the initial causal sensing LSTM layer in conjunction with the current iteration count, an intermediate causal sensing LSTM layer is obtained. This process includes: Calculate the current iteration coefficient between the current iteration number and the maximum iteration number, based on the current iteration number. By judging the current iteration coefficients, the causal structure parameters of the initial causal perception LSTM layer are frozen; Based on standard historical engineering data, the initial causal sensing LSTM layer is trained, and the temporal dynamic parameters in the initial causal sensing LSTM layer are optimized to obtain the intermediate causal sensing LSTM layer.

8. The deep learning-based power grid engineering cycle simulation method as described in claim 7, characterized in that, Optimize the temporal dynamic parameters in the initial causal sensing LSTM layer to obtain the intermediate causal sensing LSTM layer, which includes the following steps; Counterfactual time series sequences are generated based on the causal perception module. These counterfactual time series sequences are then mixed with historical engineering data to form training batches. The time series dynamic parameters are updated using a weighted sum of prediction loss and causal consistency loss. Once the convergence condition is met, the causal structure parameters and temporal dynamic parameters are solidified to obtain the intermediate causal sensing LSTM layer.

9. The deep learning-based power grid engineering cycle simulation method as described in claim 7, characterized in that, Based on historical engineering verification data, the primary two-layer network model was validated, and the project schedule deviation rate of the primary two-layer network model was obtained, specifically including: Input the verification data from the historical project verification data into the primary two-layer network model to obtain the historical project progress prediction data; The difference between the actual progress data and the predicted progress data of historical projects in the historical project verification data is calculated to obtain the project progress verification difference. Based on the ratio of the project progress verification difference to the actual historical project progress data, all project progress verification differences are integrated to obtain the project progress deviation rate of the primary two-layer network model.

10. A deep learning-based power grid engineering cycle simulation device, characterized in that, The method employing deep learning-based power grid engineering cycle simulation as described in any one of claims 1-9 includes: The data acquisition module is used to acquire engineering attribute data and engineering environment data of power grid projects; The data labeling module is used to label engineering attribute data and engineering environment data according to engineering feature relationships, so as to obtain labeled engineering attribute data and labeled engineering environment data. The progress analysis module is used to combine a pre-built two-layer network model to perform time series analysis on labeled project attribute data and labeled project environment data, and determine the project progress index vector corresponding to the project time series characteristics. The progress simulation module is used to establish a project progress simulation curve based on a preset time granularity and a project progress index vector, thereby simulating the power grid project cycle.

Images