Power transmission mechanized construction pricing feature extraction and unit price prediction method

Abstract

The application discloses a kind of power transmission mechanized construction pricing feature extraction and unit price prediction method, the present application is by collection data, establishes multi-source pricing sample index;Unified terminology, unit of measurement, price base period and space area, constructs engineering correlation atlas;From the atlas extraction features, and by quality evaluation, correlation screening and redundancy elimination form target feature set;Target feature set is input by basic unit price prediction model, mechanization correction model and space-time price correction model consisting of hierarchical prediction model, and output prediction unit price, prediction interval and feature contribution;According to actual settlement bias increment updates model parameters;The method can improve the consistency of pricing feature, unit price prediction accuracy and result traceability under the conditions of complex terrain, multiple machinery coordination and price fluctuation.

Description

Technical Field

[0001] This invention relates to the field of power transmission and transformation engineering cost data processing technology, and more specifically, to a method for extracting pricing features and predicting unit prices in mechanized construction of power transmission and transformation projects. Background Technology

[0002] With the continuous improvement of mechanization in power transmission and transformation engineering construction, operations such as foundation excavation, tower erection, conductor and ground wire laying, equipment hoisting, busbar installation, road transportation, and station renovation and expansion are gradually shifting from single manual construction to multi-mechanical collaborative construction. The type of machinery, rated capacity, shift utilization rate, access distance, number of work surfaces, power outage windows, and terrain accessibility all affect the labor, material, and machinery costs corresponding to the unit project volume.

[0003] Existing pricing methods typically rely primarily on quota items, historical comprehensive unit prices, or the experience of the compilers. For the same pricing item, the names of work processes, machinery, units of measurement, and price base periods vary across different project data. Furthermore, construction logs, shift records, market prices, and settlement results are stored in a scattered manner, making it difficult to directly form comparable training samples from historical data.

[0004] On the other hand, conventional unit price forecasting typically inputs voltage level, quantity of work, and a few environmental parameters directly into the statistical model. This fails to express the structural relationships between the project object, construction procedures, machinery combinations, resource consumption, and site constraints, and also fails to distinguish the different impacts of basic unit price, mechanized construction corrections, and spatiotemporal price fluctuations. Therefore, in scenarios such as mountainous areas, narrow station areas, multiple overlapping processes, and repeated machinery entry and exit, problems easily arise such as omitted pricing features, distorted model extrapolation, lack of interpretation of forecast results, and the inability to continuously utilize actual settlement feedback.

[0005] Therefore, there is a need for a method that can uniformly organize multi-source data of mechanized construction of power transmission and transformation, extract interpretable pricing features based on the relationship between work processes and mechanical resources, and predict the unit price of target pricing items by combining mechanized construction conditions and spatiotemporal price changes. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide a method for extracting pricing features and predicting unit prices in mechanized construction of power transmission and transformation, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] Obtain design drawings, bill of quantities, construction organization data, machinery operation data, labor, material and machinery price data, construction environment data, and historical settlement data for the target power transmission and transformation project;

[0009] A multi-source pricing sample index is established based on project identifiers, pricing sub-target identifiers, and work batch identifiers. Terminology mapping, unit conversion, price base period conversion, spatial region unification, missing value handling, and outlier verification are performed on data from each source to obtain standardized pricing samples.

[0010] Based on the standardized pricing sample, a correlation graph of engineering object—process—machinery—resource—constraint—price is constructed. The correlation graph is used to record the correlation between engineering object, construction process, construction machinery, labor and material resources, site constraints and price records.

[0011] For the target pricing item, the basic attributes of the project, the efficiency of machinery operation, the utilization rate of machinery shifts, the amortization of machinery entry and exit, the coupling of processes, the constraint of work surface, and the price characteristics of time area are extracted from the correlation graph to form a candidate pricing feature set.

[0012] The candidate pricing feature set is subjected to data integrity evaluation, cross-project stability evaluation, unit price correlation evaluation, feature redundancy elimination and monotonic consistency test to obtain the target pricing feature set.

[0013] A hierarchical unit price prediction model is trained based on the target pricing feature set and historical settlement unit prices. The hierarchical unit price prediction model includes a basic unit price prediction model, a mechanization correction model, and a spatiotemporal price correction model.

[0014] The target pricing feature set of the target power transmission and transformation project is input into the hierarchical unit price prediction model. The base unit price is multiplied by the mechanization correction coefficient and the spatiotemporal price correction coefficient, and the risk correction amount is added to obtain the predicted unit price, prediction interval, cost decomposition and feature contribution results of the target pricing sub-item.

[0015] Obtain the actual settlement unit price of the target pricing sub-item, update the sample weights, feature screening results and model parameters based on the deviation between the actual settlement unit price and the predicted unit price, and output the labor demand, material demand and machine shift demand corresponding to the predicted unit price.

[0016] Preferably, the multi-source pricing sample index uses the project identifier as the primary index, the pricing sub-target identifier as the secondary index, and the work batch identifier as the tertiary index; each standardized pricing sample records at least the voltage level, project category, construction object type, terrain category, process name, machinery combination, quantity of work, price occurrence time, price occurrence area, data source, review status, and sample quality level.

[0017] Preferably, the terminology mapping maps synonymous process names, machinery names, and material names to standard codes through a preset power transmission and transformation construction terminology dictionary; the unit of measurement conversion converts the project quantity, machinery shifts, labor man-days, and material quantity to preset standard units; the price base period conversion converts historical prices to the target price base period based on the labor price index, material price index, and machinery shift price index corresponding to the price occurrence time; and the outlier verification combines box range, median absolute deviation, and construction log evidence to determine the retention, downweighting, or removal status of outlier records.

[0018] Preferably, the association graph includes engineering object nodes, process nodes, machine nodes, resource nodes, constraint nodes, and price nodes; directed edges record the work object relationship between engineering objects and processes, the sequence relationship between processes, the execution relationship between processes and machines, the consumption relationship between processes and resources, the restriction relationship between processes and constraints, and the pricing relationship between resources and prices; attribute aggregation is performed on the multi-hop neighborhood corresponding to the target pricing item to obtain a graph structure feature containing local process chains and machine resource combination information.

[0019] Preferably, the mechanical operation efficiency characteristics are determined based on the effective completed workload, effective working time, and rated mechanical capacity; the mechanical shift utilization rate characteristics are determined based on the ratio of effective working time to occupied time; the mechanical entry and exit amortization characteristics are determined based on the ratio of mechanical entry and exit costs to the work volume that can be allocated within the same entry and exit batch; the process coupling characteristics include the statistically obtained number of shared machines, the waiting time of the preceding and following processes, and the work surface congestion coefficient determined based on the conflict time of the work surface; the work surface constraint characteristics include at least the terrain slope, transportation distance, proximity to power supply, power outage window, site accessibility, and the number of machines operating simultaneously.

[0020] Preferably, obtaining the target pricing feature set includes: removing features with a missing rate exceeding a preset missing threshold; calculating cross-project coefficient of variation and time drift index for the remaining features to obtain a stability score; calculating the mutual information and rank correlation coefficient between each feature and the settlement unit price to obtain a correlation score; retaining the feature with the highest comprehensive score for feature groups with correlation exceeding a preset redundancy threshold; using recursive features to eliminate the prediction error of the validation feature subset, and performing a monotonic consistency test on features with clear engineering mechanisms.

[0021] Preferably, the basic unit price prediction model adopts a gradient boosting decision tree model, which takes the basic attribute characteristics of the project and the standard resource price characteristics as inputs and outputs a basic unit price; the mechanization correction model takes the mechanical operation efficiency characteristics, mechanical shift utilization rate characteristics, inbound and outbound amortization characteristics, process coupling characteristics, and work surface constraint characteristics as inputs and outputs a mechanization correction coefficient; the spatiotemporal price correction model takes the target price base period, the target regional price level, and the resource price volatility as inputs and outputs a time correction coefficient and a regional correction coefficient.

[0022] Preferably, when training the hierarchical unit price prediction model, the training set, validation set, and test set are divided according to the project and the time of price occurrence to avoid adjacent work batches of the same project from entering the training set and test set at the same time; the sample training weights are determined according to the sample completeness, audit credibility, and target project similarity; the unit price prediction model is trained using robust regression loss, and the lower boundary model and upper boundary model of the prediction interval are trained using quantile regression respectively.

[0023] Preferably, the risk correction amount is obtained by weighted summation of the probability of occurrence of each risk event and the unit impact amount of the risk event; when the relative deviation of a preset number of consecutive settlement samples exceeds the drift threshold, an incremental update is triggered, and the update content includes at least the resource price index, mechanical efficiency benchmark, feature screening results, model parameters and model version number.

[0024] Preferably, the feature contribution results include the direction of increase or decrease and contribution amount of each target pricing feature to the predicted unit price; when the prediction interval width exceeds a preset confidence threshold or the similarity between the target sample and the training sample is lower than a preset similarity threshold, a low confidence label requiring manual review is generated; and a phased resource demand list and machinery combination suggestions are generated based on the decomposition results of labor, material and machinery costs corresponding to the predicted unit price.

[0025] The technical effects and advantages of this invention are as follows:

[0026] This invention unifies multi-source heterogeneous data based on a three-level index architecture, effectively solving the persistent problem of data distortion in engineering cost practice. To overcome the shortcomings of traditional flat tables that easily lose process chains and mechanical combination relationships, this solution custom-constructs an engineering object association graph, thereby structurally preserving implicit information such as shared machinery and on-site constraints. In the unit price prediction stage, complex cost changes are decoupled into three independent dimensions: basic unit price, mechanization correction, and time-based pricing. This not only accurately identifies the causes of unit price fluctuations but also, combined with robust loss functions and quantile interval prediction, significantly eliminates interference from abnormal settlement samples. Finally, through a dynamic feedback update mechanism, this method can output high-confidence feature contributions and resource requirements, significantly improving the transparency of cost auditing and the efficiency of engineering applications. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0028] Figure 2 This is a schematic diagram illustrating the data processing for price feature extraction and hierarchical unit price prediction according to the present invention. Detailed Implementation

[0029] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] like Figure 1 As shown, the first step is to obtain multi-source data for the target power transmission and transformation project. For transmission line projects, the design drawing data should at least include the line voltage level, tower type, foundation type, span, conductor specifications, terrain classification, and transportation route; for substation projects, the design drawing data should at least include the equipment type, bay type, equipment weight, installation elevation, internal roads, working space, and energized proximity conditions.

[0031] The bill of quantities data provides the names of pricing items, project characteristics, quantities, and units of measurement; the construction organization data provides construction procedures, power outage plans, work area divisions, machinery configurations, and planned construction periods; machinery operation data can be derived from shift records, equipment controller logs, positioning records, lifting records, fuel consumption records, or machinery usage certificates confirmed by the supervisor; resource price data includes labor unit prices, material market prices, machinery shift unit prices, rental prices, and machinery entry and exit costs; historical settlement data includes audited quantities, audited comprehensive unit prices, reasons for cost adjustments, and audit conclusions.

[0032] The system generates a unique project identifier for each engineering project, a pricing sub-item identifier for each list item or quota sub-item, and divides work batches according to consecutive work dates, work surfaces, and machinery combinations, generating work batch identifiers. This forms a three-level index of "engineering project - pricing sub-item - work batch". The three-level index is used to aggregate multi-source records of the same work batch and also to group by engineering project during model training to prevent data from the same project from leaking into the test set.

[0033] Each sample also records the data source and verification status. Data automatically collected on-site and consistent with the construction log is assigned a higher quality level; data confirmed by both the construction and supervision parties is assigned a lower quality level; data only supplemented by one party and lacking original documentation is assigned a lower quality level. The quality level will be converted into sample weights in subsequent model training.

[0034] The system establishes a terminology dictionary for mechanized construction of power transmission and transformation projects. The dictionary includes standard process codes, standard machinery codes, standard material codes, and a set of synonyms. For example, "rotary drilling rig hole forming," "rotary drilling hole forming," and "rotary drilling" are mapped to the same standard process code; "truck crane," "truck hoist," and "wheeled crane" are mapped to their corresponding machinery codes according to their rated lifting capacity and operation type.

[0035] For project quantities, machine shifts, and resource consumption, the system converts them according to standard units of measurement. Length is standardized to meters, area to square meters, volume to cubic meters, weight to tons, labor consumption to man-days, and machine usage to shifts or hours. If the original shift is measured in eight hours, it will be converted to a standard shift based on the effective working hours and the company's shift rules.

[0036] Historical prices are discounted based on the target price base period. Labor, material, and machinery prices are calculated using their respective price indices to avoid directly mixing prices from different periods. Let the historical price be... The historical price index is The target price base period index is Then the price is converted. Determined according to the following formula:

[0037]

[0038] For missing values, if the missing field can be derived from other records in the same batch, such as the rated capacity of machinery or standard unit of measurement, it is filled in from the machinery and equipment ledger or coding dictionary; if the missing field is a key condition affecting construction costs, such as transportation distance, effective working time or power outage window, the sample is marked as pending verification and the mean is not directly used for filling.

[0039] Outlier handling employs a combined approach of statistical detection and engineering evidence. The system calculates the median and median absolute deviation for similar engineering projects, standard procedures, and similar machine combinations, marking records exceeding preset ranges as outlier candidates. It then queries construction logs, weather records, machine failure records, and power outage change records. If genuine extreme operating conditions exist, the record is retained but its generalizability weight is reduced; if engineering evidence is lacking and the record is clearly erroneous, it is discarded.

[0040] like Figure 2As shown, the system converts standardized samples into engineering object nodes, process nodes, machinery nodes, resource nodes, constraint nodes, and price nodes. Engineering object nodes record voltage level, equipment or structure type, specifications, quantity, and spatial location; process nodes record process codes, preceding processes, succeeding processes, and work batches; machinery nodes record machinery type, rated capacity, equipment status, and leasing method; resource nodes record labor types, material specifications, and consumption quantities; constraint nodes record terrain, transportation conditions, proximity to energized areas, power outage windows, work area, and environmental conditions; and price nodes record price category, price base period, region, and data source.

[0041] Engineering object nodes are connected to process nodes through the "includes process" relationship, and process nodes form a directed process chain through the "prior to" relationship; process nodes are connected to machine nodes through the "executed by machine" relationship, to resource nodes through the "consumes resources" relationship, and to constraint nodes through the "constrained" relationship; resource nodes and machine nodes are connected to price nodes through the "corresponding price" relationship.

[0042] For the target pricing item, the system extracts a neighborhood subgraph within a preset number of hops from its corresponding process node. The first hop is used to obtain direct machinery, resources, and constraints; the second hop is used to obtain upstream and downstream processes and shared machinery; and the third hop is used to obtain prices and constraints associated with upstream and downstream processes. Specifically, a graph attention network (GAT) or a graph convolutional network (GCN) is used for neighborhood aggregation, with the aggregation layer preferably set to 2 to 3 layers to prevent excessive feature smoothing. Weighted average, maximum, minimum, and dispersion aggregation are performed on numerical attributes; embedding encoding with a dimension of 16 to 64 is performed on categorical attributes; and relational weight encoding combined with the strength of engineering constraints is performed on relational types, thereby forming a high-dimensional graph structure feature vector.

[0043] Graph structures can express relationships that are difficult to express directly in a flat table, such as "the same truck crane is shared among the three processes of equipment unloading, transfer and positioning", "power outage window causes waiting after the previous process is completed", and "the transportation distance in mountainous areas requires the cost of machinery entering and leaving the site to be amortized over a smaller amount of work".

[0044] Candidate pricing features include basic engineering attributes, mechanical operation features, organizational coupling features, site constraint features, and spatiotemporal price features. Basic engineering attributes include voltage level, engineering category, type of construction object, specifications, quantity of work, terrain category, and standard resource price.

[0045] Mechanical operating efficiency reflects the degree to which a machine's actual production capacity is utilized relative to its rated capacity. Let the... The effective workload completed by this type of machinery within the target batch is Effective working time is The rated capacity of the machine per unit time is Then the efficiency of mechanical operation for:

[0046]

[0047] Machine utilization rate reflects the proportion of effective operation during the period when machinery is occupied. Let the effective operation time be... The time the machinery occupies from its entry to its exit is Then the utilization rate U of the shift is:

[0048]

[0049] The cost of mobilizing and removing machinery needs to be converted into a unit quantity characteristic based on the apportionable quantity of work within the same batch. Let the mobilization and removal cost be... The amount of work that can be allocated within the same batch of goods entering and leaving the site is: Then the amortization characteristics of entry and exit The calculation method is as follows:

[0050]

[0051] Process coupling characteristics are used to represent the impact of shared machinery and work surface conflicts on costs. The system statistically analyzes the number of shared machines, machine changeover times, preceding and following waiting times, and work surface conflict durations between the target process and adjacent processes. Let the conflict duration be denoted as... The total duration of the target batch is α is the conflict sensitivity coefficient determined based on historical projects, then the work surface congestion coefficient... It can be represented as:

[0052]

[0053] The on-site constraints include slope, transport distance, road grade, equipment installation elevation, work surface clearance, energized safety distance, power outage window length, nighttime construction ratio, rain and snow weather ratio, and the number of machines operating simultaneously. For categorical constraints, standard coding is used; for continuous constraints, normalized values ​​and interval binning values ​​are used simultaneously to balance model fitting ability and engineering interpretability.

[0054] Spatiotemporal price characteristics include the indices of labor, major materials, and machinery shifts during the target price base period, the price ratio between the target region and the benchmark region, the volatility over recent price cycles, and the degree of resource supply tightness. For steel, conductors, insulators, fuel oil, and large machinery rental fees, which experience significant price fluctuations, independent price change characteristics are extracted separately.

[0055] After candidate features are generated, the system first calculates the missing rate of each feature. Features with missing rates exceeding a preset threshold and that cannot be reliably completed using equipment ledgers or graph relationships are removed. For the remaining features, the system calculates the coefficient of variation for different engineering projects and different price periods, and compares the distribution differences between adjacent time windows to obtain a stability score.

[0056] The correlation evaluation employs both mutual information and rank correlation coefficient. Mutual information identifies nonlinear relationships, while the rank correlation coefficient determines whether the direction of feature change aligns with the direction of unit price change. For feature groups whose absolute correlation coefficient exceeds the redundancy threshold, features with clear engineering meaning, low missing rate, high stability, and stronger correlation with the target unit price are prioritized for retention.

[0057] The system employs recursive feature elimination for validation: in each round, the set of features with the lowest contribution is removed, the validation model is retrained, and the validation error is calculated; elimination stops when feature removal leads to a significant increase in the validation error. For features with clear engineering mechanisms, such as transport distance, machinery efficiency, shift utilization, and inbound / outbound amortization, monotonic consistency checks are further performed. For example, all other things being equal, an increase in transport distance should not cause a sustained decrease in the unit price related to machinery transport, and an increase in inbound / outbound amortization should not cause a decrease in the overall unit price.

[0058] The system calculates a comprehensive score for each candidate feature. Let the stability score be... Mutual information normalization value The rank-correlation normalized value is Redundancy penalty is The overall score can then be expressed as:

[0059]

[0060] Wherein, λ1, λ2, λ3, and λ4 are non-negative weight coefficients, which can be determined by minimizing the validation set error. Features with a comprehensive score below the threshold are removed, and the remaining features form the target pricing feature set.

[0061] The preset missing threshold ranges from 10% to 30%, and in this embodiment, the preset missing threshold is 20%; the preset redundancy threshold ranges from 0.75 to 0.95, and in this embodiment, the preset redundancy threshold is 0.85.

[0062] The hierarchical unit price prediction model includes a basic unit price prediction model, a mechanization correction model, and a spatiotemporal price correction model. Before model training, the basic engineering attribute features, machinery operation efficiency features, machinery shift utilization rate features, machinery entry and exit amortization features, process coupling features, work surface constraint features, time area price features, and graph structure features of each historical pricing sample are arranged in a preset order to form an input feature vector. The audited historical settlement comprehensive unit price is converted to a unified price base period and a unified benchmark area, which serves as the supervised training label for the corresponding historical pricing sample.

[0063] The basic unit price prediction model is used to learn the relationship between the attributes of the engineering object and the standard resource price and the unit price under standard construction conditions. A gradient boosting decision tree model can be used, taking voltage level, engineering category, construction object type, specifications, standard engineering quantity, labor unit price, material benchmark price, and machinery benchmark shift price as inputs, and outputting the basic unit price. .

[0064] When training the basic unit price prediction model, samples that meet preset standard construction conditions are selected from historical pricing samples as basic training samples. The standard construction conditions include at least the following: the terrain category is within the preset benchmark category, the transportation distance is within the preset benchmark range, there is no significant conflict on the work surface, the utilization rate of machinery shifts is within the benchmark range, and there are no major power outage changes.

[0065] For the basic training samples, their historical settlement unit prices are converted according to the target price base period and benchmark region, and risk adjustment costs that can be identified separately are deducted to obtain the basic unit price training label. The basic unit price prediction model learns the unit price change relationship under standard construction conditions based on the basic attribute characteristics of the project and the standard resource price characteristics.

[0066] The mechanization correction model is used to learn the corrections made by actual mechanized construction conditions relative to standard conditions. Its inputs include machine operation efficiency, shift utilization, entry and exit amortization, number of shared machines, number of changeovers, waiting time, work surface congestion coefficient, transport distance, terrain slope, and power outage window. The output is the mechanization correction coefficient. To avoid unreasonable extrapolation, monotonically increasing constraints can be set for transportation distance, amortization of inbound and outbound costs, and conflict duration, while monotonically decreasing unit cost constraints can be set for machine efficiency and shift utilization.

[0067] The training of the mechanized correction model is performed after the training of the basic unit price prediction model is completed. For the first... A historical pricing sample is used as the basis for the basic unit price prediction model. The basic unit price prediction value is obtained by inputting the basic engineering attribute characteristics of the historical pricing sample into the basic unit price prediction model. The time correction factor is determined based on historical price indices and regional price levels. and regional correction coefficient Determine the risk adjustment amount based on the identified risk events. The first The historical settlement unit price of each historical pricing sample is recorded as follows: The corresponding mechanization correction coefficient training label Determined according to the following formula:

[0068]

[0069] The mechanical operation efficiency characteristics, mechanical shift utilization characteristics, mechanical entry and exit amortization characteristics, process coupling characteristics, and work surface constraint characteristics are used as inputs, and the training labels of the mechanization correction coefficients are used as the supervision output to train the mechanization correction model.

[0070] The mechanization correction model employs a gradient boosting decision tree model with monotonic constraints. Monotonically increasing unit cost constraints are set for transportation distance, amortization of machine entry and exit, and duration of work area conflicts, while monotonically decreasing unit cost constraints are set for machine operation efficiency and machine shift utilization rate, in order to reduce the probability of the model producing prediction results that do not conform to engineering mechanisms.

[0071] The spatiotemporal price correction model outputs time correction coefficients respectively. and regional correction coefficient The time correction factor is determined based on the labor, material, and machinery price indices of the target price base period and the model benchmark price base period; the regional correction factor is determined based on the ratio of labor, major materials, fuel, and machinery rental prices between the target region and the benchmark region.

[0072] Let the probability of occurrence of the k-th type of risk event in the risk event set be... The impact on the unit quantity of the target pricing item is Then the risk correction ΔP can be obtained by summing the probability-weighted impacts of each risk event. The final predicted unit price... Determined according to the following formula:

[0073]

[0074] The probability of occurrence of each type of risk event in the risk event set can be obtained by statistically analyzing the actual frequency of risk events in similar historical projects; the unit impact amount of the risk event can be calculated by analyzing the reasons for cost adjustments and the corresponding adjustment amounts in historical settlement data.

[0075] During model training, the model is divided into training, validation, and test sets based on project type and price occurrence time. Early projects primarily use the training set, while later projects primarily use the validation and test sets. Different batches of the same project are kept within the same dataset. This allows for simulation of the model's actual predictive ability for new projects and future price cycles.

[0076] The weight of each training sample is composed of data completeness, review credibility, and similarity to the target project. Let the three normalized scores be I, A, and M, with corresponding weight coefficients a, b, and c, then the sample weight w is:

[0077]

[0078] The point prediction model uses robust regression loss to reduce the impact of a small number of abnormal settlement samples; the interval model uses quantile regression to output lower and higher quantiles respectively, such as the 10th percentile and 90th percentile as the lower and upper boundaries of the prediction interval.

[0079] When predicting the target project, the system obtains the target pricing feature set by following the same standardization, graph construction, and feature extraction process as in the training phase. The basic unit price prediction model outputs... Mechanized correction model output Spatiotemporal price correction model output and Then, based on the risk event information, ΔP is obtained, and finally the predicted unit price and the predicted range are obtained.

[0080] The system simultaneously calculates feature contributions. For tree models, a feature contribution method based on model paths can be used to decompose the difference between the predicted unit price and the average unit price of the samples into individual target features. The output should include at least the top three features that increase the unit price, the top three features that decrease the unit price, and their respective contribution amounts. For example, transportation distance in mountainous areas may increase the unit price, higher mechanical efficiency may reduce the unit machinery cost, and smaller batch quantities may increase the amortization of inbound and outbound costs.

[0081] When the prediction interval width exceeds a preset confidence threshold, or the distance between the target sample and the training samples exceeds a threshold, the system generates a low-confidence flag. Low-confidence samples are not directly included in the automatic pricing results; instead, the compiler is shown missing conditions, similar historical items, key discrepancies, and suggested supplementary data. After supplementing the data, the target features are recalculated and predictions are made again.

[0082] The system decomposes costs based on the forecast results, breaking down the forecast unit price into labor costs, material costs, machinery costs, entry and exit amortization costs, and risk correction costs. It then generates a phased resource requirement list based on the target project volume. The machinery requirement list includes at least the machinery type, capacity level, estimated shifts, entry stage, and applicable work processes, and can be used to assist in machinery configuration and resource planning.

[0083] The aforementioned preset confidence threshold specifically refers to the difference between the upper and lower limits of the prediction interval accounting for more than 15% to 25% of the predicted unit price; the aforementioned preset similarity threshold ranges from 0.6 to 0.8.

[0084] After the target pricing item is settled, the system reads the approved quantity of work and the approved unit price, and converts them to the same price base period used in the model prediction. Let the actual settlement unit price be... Then the relative deviation D is:

[0085]

[0086] The system updates different objects according to the source of the deviation. When the deviation mainly comes from changes in labor, material, or machinery market prices, the price index and regional price ratio are updated; when the deviation mainly comes from machinery efficiency or shift utilization, the corresponding machinery efficiency benchmark is updated; when the deviation is related to unidentified work area conflicts or power outage changes, constraint nodes and relationships are added to the correlation graph; when a feature continuously loses its relevance, feature filtering is re-executed.

[0087] When the relative deviation of a preset number of consecutive settlement samples exceeds the drift threshold, an incremental model update is triggered. The incremental update retains the original model version, training sample range, parameter summary, validation error, and approval records, and generates a new version number. If the new version's error on the independent validation set is not better than the original version, the current release version will not be replaced.

[0088] After triggering the incremental update, the system first performs the same term mapping, unit of measurement conversion, price base period conversion, correlation graph construction, and target pricing feature extraction processing on the newly added settlement samples as on the historical training samples, and then adds the processed newly added settlement samples to the incremental training sample set.

[0089] The missing rate, cross-project stability score, unit price correlation score, and redundancy of each candidate pricing feature are recalculated on the incremental training sample set. When the correlation of the original target pricing feature is lower than the preset correlation threshold due to the addition of new settlement samples, or when the comprehensive score of the new candidate pricing feature is higher than that of the original target pricing feature, the target pricing feature set is redefined.

[0090] Subsequently, the basic unit price prediction model and the mechanization correction model are trained again using the original model parameters as initial parameters, or the model is retrained based on all the updated training samples. After the update, the mean absolute error, mean absolute percentage error, and prediction interval coverage of the original model and the updated model on the independent validation set are calculated respectively.

[0091] When the mean absolute error and mean absolute percentage error of the updated model are not higher than those of the original model, and the prediction interval coverage meets the preset coverage requirement, the updated model will be determined as the new release model; otherwise, the original release model will be retained, and the newly added settlement samples will continue to be saved in the sample set to be updated.

[0092] Each incremental update records the range of new samples, the target pricing feature set, model training parameters, validation metrics, model version number, update time, and review status to support the traceability of prediction results and the model update process.

[0093] The preset quantity is 3 to 10 consecutive settlement samples; the drift threshold ranges from 5% to 15%.

[0094] Taking the mechanized foundation excavation of a 220kV transmission line tower in a mountainous area as an example, the target tower foundation excavation volume is 64 cubic meters. The construction machinery consists of a 0.8 cubic meter hydraulic excavator and a dump truck. The transportation distance for the machinery to the site is 26 kilometers. The working face has a large slope, limiting the simultaneous operation of two machines. The system extracts samples with the same voltage level, similar foundation type, and similar terrain from historical projects and performs price base period conversion.

[0095] The correlation graph shows that excavators share tasks between the excavation and loading processes, and dump truck transportation is limited by the conditions of passing on mountain roads, resulting in some waiting. The system calculates the mechanical operation efficiency, shift utilization rate, entry and exit amortization, and work surface congestion coefficient, and inputs the transportation distance, slope, workload, and target price base period index into the hierarchical prediction model.

[0096] The extracted target feature set is input into the trained hierarchical prediction model. After calculation, the basic unit price prediction model outputs 356.00 yuan per cubic meter, with a mechanization correction coefficient of 1.12, a time correction coefficient of 1.03, a regional correction coefficient of 1.01, and a risk correction of 4.50 yuan per cubic meter. Therefore, the predicted unit price is approximately 419.35 yuan / cubic meter. The interval model outputs a prediction range of 398.20 yuan / cubic meter to 445.60 yuan / cubic meter.

[0097] In this example, transportation distance and amortization of entry and exit costs are the main features increasing the price, while excavator operating efficiency is the main feature decreasing the price. If the actual approved unit price is 426.80 yuan / cubic meter, the relative deviation is approximately 1.75%, and this sample can be used as a high-quality feedback sample for subsequent incremental updates. The above values ​​are only used to illustrate the data processing process of this invention and do not constitute a limitation on specific project prices.

[0098] Example 1

[0099] Taking the mechanized hoisting sub-item of a 500kV substation equipment replacement project as an example, the target equipment weight, hoisting radius, installation elevation, turning radius of roads within the substation, power outage window, and proximity to energized facilities are used as project objects and constraint node attributes. The rated lifting capacity of the truck crane, actual hoisting time, standby time, and entry / exit costs are used as mechanical nodes and price node attributes.

[0100] The system identifies the sequence of equipment unloading, in-station transfer, hoisting and positioning, adjustment and fixing, and machinery removal through the process chain, and also identifies the sharing of the same truck crane across multiple processes. If a shortened power outage window causes the crane to arrive on site earlier, the waiting time is entered into the mechanization correction model; if the same batch includes the hoisting of other equipment, the entry and exit costs are distributed among the quantities of work for multiple equipment.

[0101] In addition to providing the comprehensive unit price of the target equipment, the prediction results also output the estimated shifts for the truck crane, auxiliary lifting machinery, installation technician man-days, and material requirements for major measures. When the similarity between the target sample and the training sample is low due to the rarity of the equipment model, the system generates a low-confidence flag and prompts the user to supplement the equipment's center of gravity position, lifting lug conditions, and lifting scheme parameters before re-predicting.

[0102] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting pricing features and predicting unit prices in mechanized power transmission and transformation construction, characterized in that, include: Obtain design drawings, bill of quantities, construction organization data, machinery operation data, labor, material and machinery price data, construction environment data, and historical settlement data for the target power transmission and transformation project; A multi-source pricing sample index is established based on project identifiers, pricing sub-target identifiers, and work batch identifiers. Terminology mapping, unit conversion, price base period conversion, spatial region unification, missing value handling, and outlier verification are performed on data from each source to obtain standardized pricing samples. Based on the standardized pricing sample, a correlation graph of engineering object—process—machinery—resource—constraint—price is constructed. The correlation graph is used to record the correlation between engineering object, construction process, construction machinery, labor and material resources, site constraints and price records. For the target pricing item, the basic attributes of the project, the efficiency of machinery operation, the utilization rate of machinery shifts, the amortization of machinery entry and exit, the coupling of processes, the constraint of work surface, and the price characteristics of time area are extracted from the correlation graph to form a candidate pricing feature set. The candidate pricing feature set is subjected to data integrity evaluation, cross-project stability evaluation, unit price correlation evaluation, feature redundancy elimination and monotonic consistency test to obtain the target pricing feature set. A hierarchical unit price prediction model is trained based on the target pricing feature set and historical settlement unit prices. The hierarchical unit price prediction model includes a basic unit price prediction model, a mechanization correction model, and a spatiotemporal price correction model. The target pricing feature set of the target power transmission and transformation project is input into the hierarchical unit price prediction model. The base unit price is multiplied by the mechanization correction coefficient and the spatiotemporal price correction coefficient, and the risk correction amount is added to obtain the predicted unit price, prediction interval, cost decomposition and feature contribution results of the target pricing sub-item. Obtain the actual settlement unit price of the target pricing sub-item, update the sample weights, feature screening results and model parameters based on the deviation between the actual settlement unit price and the predicted unit price, and output the labor demand, material demand and machine shift demand corresponding to the predicted unit price.

2. The method for extracting pricing features and predicting unit prices in mechanized power transmission and transformation construction according to claim 1, characterized in that: The multi-source pricing sample index uses the project identifier as the primary index, the pricing sub-target identifier as the secondary index, and the work batch identifier as the tertiary index. Each standardized pricing sample records at least the voltage level, project category, construction object type, terrain category, process name, machinery combination, quantity, price occurrence time, price occurrence area, data source, review status, and sample quality level.

3. The method for extracting pricing features and predicting unit prices in mechanized power transmission and transformation construction according to claim 1, characterized in that: The terminology mapping maps synonymous process names, machinery names, and material names to standard codes through a pre-set power transmission and transformation construction terminology dictionary; the unit of measurement conversion converts engineering quantities, machine shifts, labor man-days, and material quantities to pre-set standard units; the price base period conversion converts historical prices to the target price base period based on the labor price index, material price index, and machine shift price index corresponding to the price occurrence time; the outlier verification combines box range, median absolute deviation, and construction log evidence to determine the retention, downweighting, or removal status of outlier records.

4. The method for extracting pricing features and predicting unit prices in mechanized construction of power transmission and transformation according to claim 1, characterized in that: The association graph includes project object nodes, process nodes, machine nodes, resource nodes, constraint nodes, and price nodes. Directed edges record the work object relationships between project objects and processes, the sequence relationships between processes, the execution relationships between processes and machines, the consumption relationships between processes and resources, the restriction relationships between processes and constraints, and the pricing relationships between resources and prices. Attribute aggregation is performed on the multi-hop neighborhoods corresponding to the target pricing item to obtain a graph structure feature containing local process chains and machine resource combination information.

5. The method for extracting pricing features and predicting unit prices in mechanized power transmission and transformation construction according to claim 1, characterized in that: The mechanical operation efficiency characteristics are determined based on the effective completed workload, effective working time, and rated mechanical capacity; the mechanical shift utilization rate characteristics are determined based on the ratio of effective working time to occupied time; the mechanical entry and exit amortization characteristics are determined based on the ratio of mechanical entry and exit costs to the work volume that can be allocated within the same entry and exit batch; the process coupling characteristics include the statistically obtained number of shared machines, the waiting time of the preceding and following processes, and the work surface congestion coefficient determined based on the conflict time of the work surface; the work surface constraint characteristics include at least the terrain slope, transportation distance, proximity to power supply, power outage window, site accessibility, and the number of machines operating simultaneously.

6. The method for extracting pricing features and predicting unit prices in mechanized construction of power transmission and transformation according to claim 1, characterized in that: The acquisition of the target pricing feature set includes: removing features with a missing rate exceeding a preset missing threshold; calculating cross-project coefficient of variation and time drift index for the remaining features to obtain a stability score; calculating the mutual information and rank correlation coefficient between each feature and the settlement unit price to obtain a correlation score; retaining the feature with the highest comprehensive score for feature groups with correlation exceeding a preset redundancy threshold; using recursive features to eliminate the prediction error of the validation feature subset, and performing a monotonic consistency test on features with clear engineering mechanisms.

7. The method for extracting pricing features and predicting unit prices in mechanized construction of power transmission and transformation according to claim 1, characterized in that: The basic unit price prediction model adopts a gradient boosting decision tree model, which takes the basic attribute characteristics of the project and the standard resource price characteristics as inputs and outputs a basic unit price; the mechanization correction model takes the mechanical operation efficiency characteristics, mechanical shift utilization rate characteristics, entry and exit amortization characteristics, process coupling characteristics, and work surface constraint characteristics as inputs and outputs a mechanization correction coefficient; the spatiotemporal price correction model takes the target price base period, the target regional price level, and the resource price volatility as inputs and outputs a time correction coefficient and a regional correction coefficient.

8. The method for extracting pricing features and predicting unit prices in mechanized construction of power transmission and transformation according to claim 1, characterized in that: When training the hierarchical unit price prediction model, the training set, validation set, and test set are divided according to the project and the time of price occurrence to avoid adjacent batches of the same project from entering the training set and test set at the same time; the sample training weights are determined according to the sample completeness, audit credibility, and target project similarity; the robust regression loss is used to train the unit price point prediction model, and quantile regression is used to train the lower boundary model and upper boundary model of the prediction interval respectively.

9. The method for extracting pricing features and predicting unit prices in mechanized construction of power transmission and transformation according to claim 1, characterized in that: The risk correction amount is obtained by weighted summation of the probability of occurrence of each risk event and the unit impact of the risk event; When the relative deviation of a preset number of consecutive settlement samples exceeds the drift threshold, an incremental update is triggered. The update includes the resource price index, mechanical efficiency benchmark, feature selection results, model parameters, and model version number.

10. The method for extracting pricing features and predicting unit prices in mechanized construction of power transmission and transformation according to claim 1, characterized in that: The feature contribution results include the direction and amount of increase or decrease of each target pricing feature on the predicted unit price; when the prediction interval width exceeds the preset confidence threshold or the similarity between the target sample and the training sample is lower than the preset similarity threshold, a low confidence label requiring manual review is generated; based on the decomposition results of labor, material and machinery costs corresponding to the predicted unit price, a phased resource demand list and machinery combination suggestions are generated.

Images