A shipyard air compression station intelligent scheduling system based on a big data algorithm

Abstract

The application provides a shipyard air compression station intelligent scheduling system based on a big data algorithm, and relates to the technical field of industrial intelligence optimization. The system comprises an initial input data module, an evaluation system module, an experience pool module, a multi-dimensional state input module, a unit scheduling model module, a boundary condition fusion module and an operation feedback module. The system uses a feedforward neural network model and a closed-loop feedback mechanism to realize dynamic optimization scheduling of air compression units. Through multi-source data acquisition and quality evaluation, the system constructs a dynamic experience pool, combines mechanism flow calculation and multi-dimensional state integration, outputs qualitative scheduling suggestions and converts them into quantitative operation instructions, and continuously optimizes scheduling strategies. Through the big data algorithm and modular design, the application significantly improves the accuracy and energy efficiency of air compression unit scheduling, reduces energy consumption and equipment wear, enhances the adaptability of the system to dynamic working conditions, and ensures the stability of production and the maximization of energy utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent optimization technology, specifically to an intelligent scheduling system for shipyard air compressor stations based on big data algorithms. Background Technology

[0002] Industrial energy efficiency management is a crucial means of achieving green and low-carbon development. For shipbuilding companies, power gas energy is a significant component of their energy consumption. Air compressors alone account for approximately 30% to 40% of a shipbuilding company's total annual power generation, and operating costs account for up to 70% of their total lifecycle costs. While shipyard pneumatic systems and air compressor equipment are designed and selected based on maximum production demand, actual usage is subject to significant fluctuations in air consumption due to changes in shipyard production conditions. This necessitates frequent and substantial adjustments to the air output of shipyard air compressor units. Traditional air compressor station scheduling strategies heavily rely on manual experience or fixed rules, making it difficult to address dynamic fluctuations in air demand (such as intermittent air usage during processes like welding and sandblasting). This often leads to unstable supply pressure, low load rates, and high energy consumption.

[0003] Currently, most enterprises in my country still manage air compressor operation using traditional manual scheduling methods. Scheduling personnel primarily rely on on-site observation and experience to adjust compressed air supply, lacking scientific theoretical guidance. Therefore, research on intelligent scheduling systems for shipyard air compressor stations is of significant practical importance. Summary of the Invention

[0004] In view of the technical problems of air compressor unit scheduling relying on manual experience or fixed rules in the existing technology, which has the problems of slow response and high energy consumption, this invention proposes an intelligent scheduling system for shipyard air compressor stations based on big data algorithms.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: an intelligent scheduling system for shipyard air compressor stations based on big data algorithms, the system comprising: The initial input data module is used to collect and store the original operating data of the air compressor station and the output data of the load prediction model, which serve as the initial input data for big data algorithms. The evaluation system module evaluates the quality of the initial input data according to preset rules, selects high-quality samples to form a training set and outputs them to the experience pool module. The experience pool module stores high-quality training sets, supports dynamic updates, and provides training samples for the unit scheduling model module. The multi-dimensional state input module integrates multi-dimensional state variables to form a real-time multi-dimensional state vector and outputs it to the unit scheduling model module. The unit scheduling model module adopts a feedforward neural network structure and outputs qualitative scheduling suggestions L1 to the boundary condition fusion module based on the real-time multidimensional state vector; The boundary condition fusion module integrates the air compressor's operating constraints, transforms the qualitative scheduling suggestion L1 into an operable quantitative scheduling suggestion L2, and outputs it to the operation feedback module. The operation feedback module records staff operations and sends them back to the system, forming a closed-loop data flow to support strategy optimization.

[0006] Compared with the prior art, the beneficial effects of the present invention are: By constructing an evaluation system module to assess the quality of the initial input data, selecting data samples with high stability and strong representativeness, and training them in conjunction with a feedforward neural network model, the scheduling matching rate of air compressor units under low-pressure conditions was significantly improved (the matching rate of Type I air compressors increased by 5.15%, and Type II by 2.32%), the complete matching rate increased by 4.26%, and operational deviations were significantly reduced.

[0007] The evaluation system uses a quantitative scoring mechanism to screen the initial input data for quality. Based on the dual indicators of compliance and volatility of pressure range, it accurately eliminates low-quality data. This allows the high-quality data samples stored in the experience pool to improve the operational quality score of the scheduling model by 1.7 points under low-pressure conditions (the overall average score increases from 73.15 to 74.85). At the same time, it significantly improves the model training effect, ensures the stability and reliability of scheduling suggestions, and provides a solid data foundation for intelligent optimization scheduling of air compressor units.

[0008] The unit scheduling model integrates multi-dimensional state variables (such as real-time flow, predicted flow, mechanistic flow, and pressure) through a feedforward neural network architecture. Combined with a multi-task learning framework, it achieves accurate joint prediction of the number of Type I and Type II air compressors loaded, improving the scheduling matching rate by more than 5% under low-pressure conditions. At the same time, through boundary condition fusion, it transforms qualitative suggestions into quantitative operation instructions that meet the equipment start-up and shutdown constraints. Finally, under the closed-loop feedback mechanism, the deviation between theoretical flow and actual operating conditions is reduced to within 850 m³ / h, significantly improving the energy efficiency and stability of air compressor unit operation.

[0009] Furthermore, this invention employs a modular design to support strong compatibility, adapting to air compressor unit configurations in shipyards of varying sizes. The closed-loop feedback mechanism continuously optimizes the scheduling strategy through an operational feedback module, enabling the system to maintain efficient operation even under dynamic fluctuations in intermittent gas-using processes such as welding and sandblasting, demonstrating broad potential for widespread adoption.

[0010] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0011] Figure 1This is a block diagram of a shipyard air compressor station intelligent scheduling system based on big data algorithms according to the present invention; Figure 2 This is a flowchart of the evaluation system according to the present invention; Figure 3 This is a scatter plot of the original data distribution according to the present invention; Figure 4 This is a scatter plot of the data distribution after evaluation according to the present invention; Figure 5 This is a schematic diagram of the neural network structure of the unit scheduling model according to the present invention; Figure 6 This is a training graph based on the training set of the present invention, which is the total data. Figure 7 This is a training graph based on the training set of the present invention, which serves as an experience pool. Figure 8 This is the result of the June deployment experiment of the intelligent scheduling system according to the present invention. Detailed Implementation

[0012] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings, so as to more clearly understand the purpose, features and advantages of this invention. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of this invention, but are only for illustrating the essential spirit of the technical solutions of this invention. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0013] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0014] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0015] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to mean “and / or” unless otherwise expressly stated herein.

[0016] In the following description, in order to clearly demonstrate the structure and working method of the present invention, a number of directional terms will be used. However, terms such as "front", "back", "left", "right", "outside", "inside", "outward", "inward", "up", and "down" should be understood as convenient terms and not as limiting terms.

[0017] The implementation details of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following content is only for the convenience of understanding the implementation details and is not necessary for implementing this solution.

[0018] As a core energy supply network, the pneumatic system of a shipyard directly impacts the plant's energy consumption and production stability through the efficient operation of its air compressor units. However, traditional scheduling methods suffer from two major contradictions: First, dynamic supply and demand imbalances exist. Shipyard air demand exhibits strong fluctuations (such as intermittent air consumption during processes like welding and sandblasting), while air compressor units are often scheduled using a crude method of "fixed number of units + experience-based start-stop," resulting in large fluctuations in pipeline pressure. This leads to decreased efficiency of pneumatic tools (such as uneven sandblasting particle size) and energy waste. Second, equipment health and grid impacts are significant. Frequent start-stops of air compressors can cause surge phenomena, accelerating mechanical wear and impacting the power grid, increasing maintenance costs and the risk of power outages.

[0019] Example 1 To address the aforementioned issues, this invention proposes an intelligent scheduling system for shipyard air compressor stations based on big data algorithms. This system integrates multi-source data acquisition and processing, intelligent data quality assessment and dynamic experience pool construction, precise calculation of mechanistic flow and multi-dimensional state integration, and uses a feedforward neural network model to output qualitative scheduling suggestions, which are then transformed into quantitative operation instructions based on boundary conditions. Simultaneously, a closed-loop feedback mechanism is used to continuously optimize the scheduling strategy, thereby achieving efficient and accurate operation and energy efficiency improvement of the air compressor station. The system features modular deployment, strong compatibility, and broad application value.

[0020] Specifically, such as Figure 1 As shown, the intelligent scheduling system for shipyard air compressor stations based on big data algorithms provided in this embodiment includes seven modules: an initial input data module, an evaluation system module, an experience pool module, a multi-dimensional status input module, a unit scheduling model module, a boundary condition fusion module, and an operation feedback module. Initial Input Data Module: This module includes collected and stored raw operating data of the air compressor station and output data of the load forecasting model. The raw operating data includes various types of data such as the operating status of the air compressor, and the flow and pressure of the main and branch pipes. This raw operating data is unfiltered and retains all original information from the actual operation process. The raw operating data and the output data of the load forecasting model together serve as the initial input data for the big data algorithm.

[0021] Evaluation system module: Based on preset rules (such as pressure range compliance, volatility and other indicators), the initial input data received is evaluated for quality, and data samples with high stability and strong representativeness are selected to form a high-quality training dataset. The high-quality training dataset is then output to the experience pool module.

[0022] Experience Pool Module: Used to store high-quality training datasets selected by the evaluation system module, supporting dynamic updates and expansion, providing continuously optimized training samples for the unit scheduling model module, and supporting the iteration and performance improvement of the unit scheduling model.

[0023] The multi-dimensional state input module includes a mechanism flow calculation module. This module calculates the mechanism flow using a professional mechanism fitting curve and combines it with real-time pressure data to form an estimated mechanism flow value that reflects the actual operating state of the system. This estimated mechanism flow value is then output to the multi-dimensional state input module as part of the input for training the unit scheduling model.

[0024] The multidimensional state input module integrates multidimensional state variables such as the mechanism flow estimate, the actual flow of the main pipe and branch pipes, the predicted flow of the main pipe, and the current pressure to form a real-time multidimensional state vector, which is output to the unit scheduling model module for training, providing a comprehensive basis for scheduling decisions.

[0025] Unit scheduling model module: It adopts a feedforward neural network structure, based on the real-time input multi-dimensional state vector, and outputs qualitative scheduling suggestion L1 to the boundary condition fusion module. This L1 is the total number of units to be loaded for each type I and type II air compressor.

[0026] Boundary condition fusion module: Integrates the air compressor operation constraints of the air compressor station (such as the daily start-stop limit, monthly operation balance requirements and start-stop time interval limit), transforms the qualitative scheduling suggestion L1 into an operable quantitative scheduling suggestion L2, and outputs the quantitative scheduling suggestion L2 to the operation feedback module. This L2 clarifies the specific units and actions that need to be operated in the Type I and Type II air compressors.

[0027] Operation feedback module: Records the operations performed by staff based on quantitative scheduling suggestions L2 and the actual situation on site, and sends the operation records back to the system to form a closed-loop data flow, supporting the iterative optimization of scheduling strategies.

[0028] Through data flow and functional collaboration among the aforementioned modules, the shipyard air compressor station intelligent scheduling system based on big data algorithms has achieved full automation and intelligence from data collection, quality assessment, model training to scheduling suggestion output.

[0029] In some embodiments, the aforementioned evaluation system module is one of the core modules of the air compressor station intelligent dispatching system. Its core function is to comprehensively evaluate the actual effects of historical dispatching operations through a quantitative scoring mechanism, and to select high-quality data to optimize the training effect of the unit dispatching model. The scoring process is based on dual indicators of pressure range compliance and volatility, and achieves accurate identification of data quality through dynamic weight allocation, such as... Figure 2 As shown, the specific process is as follows: (1) Input data acquisition and preprocessing The system receives the following data as the basis for scoring: Current status data: including the current system pressure value (Unit: bar), real-time flow rate of main pipe and branch pipes (unit: m) 3 / h); Historical moment status data: Pressure value at the previous moment (Unit: bar) Flow rate of main pipe and branch pipes at the previous moment (unit: m) 3 ( / h) and the scheduling scheme executed in the previous time step. .

[0030] (2) Interval compliance scoring mechanism This indicator is used to measure whether the current pressure is stable within the target pressure range. The scoring rules are as follows: Assess current stress levels Whether the pressure (in bar) is within the target pressure range is assessed, and points are deducted linearly based on the degree of deviation, according to the following rules: For example, when the pressure value is 5.0 bar, its interval compliance score is 42.5, and the absolute deviation from the target center value of 5.5 bar is 0.5 bar, which accounts for 50% of the entire allowable interval width (1.0 bar).

[0031] (3) Volatility scoring mechanism This section combines pressure and flow rate changes and scores them based on a pre-defined piecewise function. Different flow rate change ranges correspond to different pressure fluctuation tolerance thresholds and deduction functions, as detailed below: The volatility scoring mechanism calculates the system's stability score by combining pressure variation (ΔP, in bar) and flow variation (ΔF, in m³ / h). ΔP is defined as the absolute difference between the current pressure and the previous pressure, while ΔF is defined as the absolute difference between the current flow and the previous flow. This scoring mechanism is based on a piecewise function. When ΔP ≤ 0.1, a full score of 50 is awarded, indicating minimal system pressure fluctuation. When flow variation is small (e.g., ΔF < 10000), if 0.1 < ΔP < 0.2, the score is calculated according to the second formula. For example, with ΔP = 0.15 and ΔF = 8000, the deduction is based on the portion exceeding 0.1 bar, i.e., an absolute deviation of 0.05 bar, which occupies 50% of the maximum allowable pressure fluctuation range (0.2 - 0.1 = 0.1 bar) within the current flow range. =50-50^(0.15-0.1) / 0.1=42.9 points, indicating that the system is in a low-fluctuation state and has good stability. As the flow rate changes more, the tolerance range for pressure fluctuations also gradually widens. For any situation that does not meet the above conditions, the score is directly recorded as 0 points, indicating that the system volatility is too high. Overall, this mechanism reflects the impact of flow rate changes on system stability by dynamically adjusting the pressure fluctuation threshold, ensuring that the scoring can both capture subtle fluctuations and adapt to different operating conditions.

[0032] (4) The total score is obtained by adding the above two parts: like Then determine the scheduling scheme. If effective, it will be stored in the experience pool for subsequent training of the unit scheduling model and optimization of scheduling strategies.

[0033] The data after evaluation of the system is entered into the experience pool, and the storage format is shown in Table 1: Table 1. Experience Pool Data Format If we assume that we only need to consider flow rate differences greater than 6000 m³ / min within 15 minutes... 3 Scheduling is performed based on the / h operating condition. The original data scatter plot is shown below. Figure 3 As shown; after evaluation by the system, the scatter plot of the data entering the experience pool is as follows. Figure 4 As shown, the evaluation system yields high-quality and effective data as an experience pool. This high-quality data is beneficial for optimizing the training effect of the unit scheduling model, improving the accuracy and reliability of future scheduling decisions, and supporting the continuous improvement of system operation strategies. Future plans include further enhancing the diversity and time-series coverage of the experience pool data, introducing raw operational data with a wider time span and more comprehensive operating conditions, and optimizing the evaluation rules in conjunction with professional advice.

[0034] In some embodiments, training and test sets are divided into time series based on the original operational data to ensure the timeliness and accuracy of the intelligent scheduling system evaluation. The evaluation system module performs quality screening on the training set, and using the aforementioned comprehensive scoring mechanism, quantitatively scores each data record based on indicators such as stress interval compliance and system volatility, setting a threshold (…). High-quality data is selected and stored in a dedicated experience pool to provide excellent samples for training the unit scheduling model; at the same time, the training set retains complete data records for subsequent comparative analysis.

[0035] Based on the selected high-quality data, a unit scheduling model was constructed. A feedforward neural network architecture was adopted to design the unit scheduling model, which can effectively handle multi-dimensional state input data. A rigorous checking mechanism was implemented during the training process of the unit scheduling model, including setting clear training indicators and convergence criteria, real-time monitoring of the loss function changes during training, using cross-validation to prevent overfitting, and periodically saving training snapshots to facilitate backtracking and comparison of model performance at different stages. Through iterative optimization, the unit scheduling model can accurately learn the system's operating rules and output reasonable scheduling suggestions.

[0036] Specifically, this intelligent scheduling system employs a deep learning model based on a feedforward neural network for air compressor scheduling prediction. The network structure comprises one input layer, three hidden layers, and two independent output layers, forming a multi-task learning framework. This design enables joint prediction of the loading quantity of both Type I and Type II air compressors.

[0037] (1) Input layer design The network input is a 1×4 dimensional feature vector, containing the following four key features: Flow: Current measured traffic volume; Future Flow: Predicts traffic flow for the next 15 minutes; Pressure: Current system pressure value; Mechanism_Flow: The flow value calculated using a specialized mechanistic model.

[0038] (2) Hidden layer configuration The network contains three fully connected hidden layers and uses the ReLU activation function. Specific parameter configurations are as follows: Figure 5As shown, the network structure is a multi-layer feedforward neural network with a fully connected layer design. It processes data step by step starting from the input layer: the input layer (fc1) receives 4-dimensional features, which are expanded to 1024 dimensions through the first fully connected layer (fc2), and the ReLU activation function is applied to introduce non-linearity; subsequently, the fc3 layer compresses the dimension to 512 dimensions, and the fc4 layer further compresses it to 256 dimensions. Both of these layers retain the ReLU activation to ensure the expressive power of the model; finally, the network structure is divided into two independent output layers.

[0039] (3) Output layer design The network structure employs two independent output layers to predict the load quantities of the two types of air compressors, respectively: num_big output layer (Type I air compressor): Weighting dimensions: 9 × 128; Bias dimension: 9; Output a 1×9 vector representing the probability distribution of the number of large air compressors loaded from 0 to 8.

[0040] num_small output layer (Type II air compressor): Weight dimension: 2 × 128; Bias dimension: 2; Output a 1×2 vector representing the probability distribution of the number of small air compressors loaded (0-1).

[0041] Two independent output layers generate 9-dimensional results (representing 0-8 loaded Type I air compressors) and 2-dimensional results (representing 0-1 loaded Type II air compressors), respectively. The two output tasks share all features from layers fc1 to fc4. This structure allows the model to learn common features from both tasks while also enabling each output layer to learn unique feature transformations, avoiding interference between tasks and facilitating flexible adjustment of the output dimension and loss function for individual tasks.

[0042] This design adopts a multi-task learning architecture with a shared bottom layer and independent outputs, which has the following technical advantages: the two output tasks share all network layers from the input layer to the fullc4 layer, realizing the joint learning and sharing of feature representations; each prediction task has an independent output layer, which can learn its own unique feature transformations and avoid mutual interference between tasks; it is convenient to adjust the output dimension or loss function weight of a single task according to actual needs; the shared feature extraction layer reduces the total number of parameters and improves training efficiency and generalization ability.

[0043] Furthermore, the intelligent scheduling system employs a deep learning model based on a feedforward neural network for air compressor scheduling prediction, achieving joint prediction of the loading quantity of Type I and Type II air compressors. The network hyperparameter settings for the unit scheduling model in the intelligent scheduling system include: (1) Learning rate setting The learning rate is a core hyperparameter controlling the step size of parameter updates during deep learning model training, directly affecting model convergence. A suitable learning rate can effectively balance training speed and convergence stability. Based on practical experience, the common learning rate range is usually set between 0.0001 and 0.1. For the Adam optimizer (adaptive moment estimator) used in this invention, its dynamic learning rate adjustment feature allows for relatively flexible initial value setting. After multiple experimental verifications, 0.0004 was ultimately chosen as the initial learning rate for model training. This choice is based on the following considerations: providing stable convergence characteristics; achieving a good balance between training efficiency and model performance; and being suitable for the characteristics of the medium-sized dataset in this project.

[0044] (2) Batch size settings Batch size is another important hyperparameter, directly affecting the model's training performance and computational efficiency. Based on the training set size (over 100,000 samples, considered a medium-sized dataset), this invention adopts the following setting strategy: Referring to medium-sized datasets (10k-1M samples), official documentation and academic papers of mainstream deep learning frameworks (such as TensorFlow / PyTorch) typically recommend batch sizes ranging from 64 to 1024. After multiple rounds of experimental comparisons, a batch size of 256 was ultimately determined. This choice was based on the following considerations: fully utilizing the parallel computing power of GPUs to improve training efficiency; maintaining sufficient gradient stability to avoid drastic fluctuations during training; and achieving the optimal balance between memory constraints and training stability.

[0045] The batch size of the network hyperparameter is set to 256, which, together with the learning rate of 0.0004, achieves gradient stability. When the batch size increases by a factor of k, the gradient variance decreases by a factor of k, requiring synchronous adjustment of the learning rate to maintain the signal-to-noise ratio. This invention increases the learning rate accordingly. This is to maintain the same "signal-to-noise ratio" during training.

[0046] Furthermore, during the training of the unit scheduling model, the cross-entropy loss function is used as the optimization objective. This loss function is widely used in classification tasks and can effectively measure the difference between the probability distribution predicted by the model and the true labels. Loss values ​​are calculated separately for the prediction tasks of the two types of air compressors. Considering the difference in importance between the two types of air compressors in the system, a weighted summation loss function is used: Total loss function value = × Loss value of small air compressor + Loss value of large air compressor By adjusting the weight parameters This can balance the impact of the two types of prediction tasks on the overall optimization process, and is currently set to... .

[0047] For example, the training process was monitored on two different training sets, with a data time range from March 1 to May 12, 2025, respectively, on the original dataset and the dataset selected by the evaluation system. Figure 6 and Figure 7 As shown, the continuous decrease in the loss function proves that the model parameters are being updated in the right direction; the convergence of the loss value indicates that the model has learned the effective features in the data and reached a stable state; the training strategy is reasonable, and the selected hyperparameters such as learning rate and batch size are appropriately configured.

[0048] The good convergence properties of the loss function prove the effectiveness of the designed network structure and training method, and provide a strong guarantee for the stability of the model in practical applications.

[0049] This embodiment proposes an intelligent scheduling system for shipyard air compressor stations based on big data algorithms. By integrating an initial input data module, an evaluation system module, an experience pool module, a multi-dimensional state input module, a unit scheduling model module, a boundary condition fusion module, and an operation feedback module, and utilizing a feedforward neural network model and a closed-loop feedback mechanism, the system achieves dynamic optimization scheduling of air compressor units. Through multi-source data acquisition and quality assessment, the system constructs a dynamic experience pool, combines mechanistic flow calculation with multi-dimensional state integration, outputs qualitative scheduling suggestions, and transforms them into quantitative operation instructions, continuously optimizing the scheduling strategy. This embodiment, through big data algorithms and modular design, significantly improves the accuracy and energy efficiency of air compressor unit scheduling, reduces energy consumption and equipment wear, and enhances the system's adaptability to dynamic operating conditions, ensuring maximum production stability and energy utilization efficiency.

[0050] Example 2 This embodiment aims to systematically verify the optimization effect of the evaluation system on scheduling performance. By comparing and analyzing the scheduling performance of the unit scheduling model under different data training conditions, the effectiveness of the data quality screening mechanism of the evaluation system is evaluated. The following four key parameters are used as inputs to the unit scheduling model: Current traffic: The traffic volume value monitored by the system in real time; Future flow: The predicted flow value for a future period; Mechanistic flow: Flow estimates calculated based on specialized mechanistic models; Current pressure: Real-time pressure monitoring data of the system.

[0051] To verify the optimization effect of the evaluation system on the air compressor scheduling of the air compressor station, this embodiment designs a comparative experiment, using the following two sets of training data to train the unit scheduling model: Group 1: Use the complete raw running data from March 1, 2025 to May 12, 2025 as the training set; The second group: Data selected by the evaluation system within the same time period (Score ≥ 85 or Stress P ≥ 650 kPa) was used as the training set.

[0052] The experiment comprehensively evaluated the performance of the two models under overall operating conditions and low-pressure operating conditions from three dimensions: matching rate, average deviation, and operational quality. The results are shown in Tables 2 and 3. Table 2 Test Results (Overall Operating Conditions) Table 3 Test Results (Low-voltage Condition) The focus is on low-pressure operating conditions. Under low-pressure conditions, although the matching rates of both models remain at a high level, there are significant differences in operational quality: For Type I air compressors: the average score for the matched portion of the first group is 62.2, while the average score for the unmatched portion is only 21.5; for the second group, the average score for the matched portion is 95.2, and the average score for the unmatched portion also reaches 94.5. A similar trend is observed for Type II air compressors, with the second group maintaining a high level of operational quality in both matched and unmatched conditions.

[0053] The results of this experiment show that: the high-quality training data selected by the evaluation system (i.e., the second group) significantly improves the quality of scheduling operations under low-pressure conditions; model robustness: both groups of models maintain extremely high matching rates and low prediction bias under different operating conditions; the evaluation system's screening mechanism effectively identifies high-quality operational experience, especially showing significant advantages under critical low-pressure conditions.

[0054] Although the second set of model training data was smaller in scale, its quality was significantly improved. While maintaining a high matching rate, it greatly improved the actual effect of scheduling operations, verifying the effectiveness and practical value of the evaluation system in optimizing air compressor scheduling.

[0055] This embodiment verifies the optimization effect of the evaluation system on air compressor scheduling through comparative experiments. The unit scheduling model was trained using both the original data and the high-quality data selected by the evaluation system. Experimental results show that the data selected by the evaluation system significantly improved the scheduling operation quality under low-pressure conditions. The average score of the matched portion increased from 62.2 to 95.2 (Type I air compressor), and the average score of the unmatched portion increased from 21.5 to 94.5, verifying the optimization effect of the data quality selection mechanism on model performance. Quantitative comparison demonstrates that the high-quality data selected by the evaluation system can significantly improve the operational accuracy and stability of the scheduling model under key operating conditions (such as low-pressure conditions), providing effective data support for model training.

[0056] Example 3 This embodiment verifies the intelligent scheduling system for air compressors based on Embodiment 2, and deploys it as follows: Figure 8 As shown: (1) Initial deployment The scheduling model was first deployed in the field in June 2025. Preliminary operational data shows that although the model's output recommendations have a high matching rate with historical operations, there is a certain deviation between the theoretical flow corresponding to the recommended scheduling schemes and the mechanistic flow calculated based on field pressure, with an average deviation of 2290 m³. 3 This phenomenon indicates that the model needs further optimization to improve its matching with actual working conditions.

[0057] (2) Optimized deployment (September 2025) Optimization Logic 1: Evaluate and Verify System Effectiveness By setting up comparative experiments, the differences in model training performance between data filtered by the evaluation system and unfiltered raw data were analyzed, thereby verifying the effectiveness of the evaluation system in improving data quality and model performance. By comparing controlled variables, the contribution of the evaluation system to subsequent optimization was clarified.

[0058] Optimization Logic 2: Scheduling Model Matching Quality Analysis A quantitative evaluation method was used to calculate the average quality score of the parts of the model output that matched and did not match historical operations. By analyzing the score differences between these two parts, the scheduling model was verified to ensure that it can accurately identify and match high-quality operation records and provide better scheduling suggestions for low-quality operations, thus demonstrating the model's optimization capability.

[0059] Optimization Logic 3: Input Data Consistency Verification The data input process was confirmed to ensure consistency between the field-collected data and the data processed by the algorithm. The verification results showed that the total pipe pressure data collected on-site was consistent with the pressure data obtained by the algorithm. However, there were some differences in the total pipe flow rate data, with the difference in average flow rates remaining ≤10 m³. 3 The discrepancy was caused by a one-second difference between the on-site data retrieval logic and the algorithm program, which has now been corrected.

[0060] This study compares and evaluates the performance of the "simultaneous matching" concept (i.e., matching both large and small air compressors) in an air compressor scheduling system under low-pressure conditions (480kPa-620kPa). The optimization effect is analyzed by comparing the operating data of the system before and after its application.

[0061] (1) Before applying the evaluation system (00:00 on August 23, 2025 - 13:46 on August 28, 2025): Matching rate analysis: The matching rate of large air compressors is 77.06%, with an average deviation of 0.26; the matching rate of small air compressors is 73.49%, with an average deviation of 0.27. Full matching rate: The proportion of both large and small air compressors being fully matched is 65.65%; Quality score: The overall average score under low-pressure conditions is 73.15 points, with an average score of 71.49 points when matched and an average score of 76.30 points when mismatched.

[0062] (2) After applying the evaluation system (September 18-22, 2025): Matching rate analysis: The matching rate of large air compressors increased to 82.21% (an increase of 5.15 percentage points), with the average deviation remaining at 0.26; the matching rate of small air compressors increased to 75.81% (an increase of 2.32 percentage points), with the average deviation decreasing to 0.24. Full match rate: The full match rate increased to 69.91%, an increase of 4.26 percentage points. Quality rating: The overall average score for low-pressure conditions improved to 74.85 points, the average score for matching conditions improved to 75.62 points, and the average score for mismatched conditions was 73.05 points.

[0063] (3) Performance Improvement Analysis After the evaluation system was implemented, the matching rate of both large and small air compressors improved, with a more significant increase in the matching rate of large air compressors. The increased proportion of perfect matching indicates that the system has improved its ability to coordinate the collaborative work of large and small air compressors.

[0064] Comparing the data from the two time periods revealed that after applying the evaluation system, the average score for matching increased from 71.49 to 75.62, while the average score for non-matching decreased, indicating that the system is more accurate in identifying high-quality operation patterns.

[0065] The application of the evaluation system has improved the scheduling performance of air compressors under low-pressure conditions, achieving substantial improvements in both matching accuracy and scheduling quality. Further optimization of scheduling strategies under mismatch conditions can be carried out to continuously improve the overall performance of the model.

[0066] This third embodiment, based on the second embodiment, demonstrates the on-site deployment and verification of the intelligent air compressor scheduling system. Initial deployment revealed a deviation between the model's output suggestions and actual operating conditions (average deviation 2290 m³ / h). After optimization (including system effectiveness verification, scheduling model matching quality analysis, and input data consistency verification), the matching rate for large air compressors increased by 5.15 percentage points, for small air compressors by 2.32 percentage points, and the complete matching rate increased by 4.26%. The overall scheduling quality score under low-pressure conditions improved from 73.15 to 74.85, confirming the system's adaptive optimization effect under dynamic operating conditions. Through on-site deployment and verification of the system's actual optimization capabilities, the source of model deviation was revealed, and the data input logic was corrected, ultimately achieving a substantial improvement in matching rate and scheduling quality, proving that the system possesses adaptive optimization capabilities under dynamic operating conditions.

[0067] The implementation process of the intelligent scheduling system for shipyard air compressor stations based on big data algorithms provided by this invention includes modules such as initial input data collection and quality assessment, dynamic experience pool construction, multi-dimensional state input integration, feedforward neural network model training, boundary condition constraint fusion, and closed-loop feedback optimization. High-quality data is screened through a quantitative scoring mechanism, and combined with mechanistic flow calculation and real-time pressure monitoring, the system outputs qualitative scheduling suggestions and transforms them into quantitative operation instructions. Finally, the model's matching accuracy and energy efficiency improvement effect under dynamic operating conditions are verified through on-site deployment. This invention patent, by integrating big data algorithms and multi-module collaborative optimization mechanisms, realizes intelligent dynamic scheduling of shipyard air compressor units, significantly improving scheduling accuracy and energy utilization efficiency.

[0068] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the invention by those skilled in the art without departing from the spirit and essence of the invention. Such modifications or substitutions should all fall within the scope of the invention, or any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the invention should be covered within the protection scope of the invention. Therefore, the protection scope of the invention should be determined by the scope of the claims.

Claims

1. A shipyard air compressor station intelligent scheduling system based on big data algorithms, characterized in that, The system includes: The initial input data module is used to collect and store the original operating data of the air compressor station and the output data of the load prediction model, which serve as the initial input data for big data algorithms. The evaluation system module evaluates the quality of the initial input data according to preset rules, selects high-quality samples to form a training set and outputs them to the experience pool module. The experience pool module stores high-quality training sets, supports dynamic updates, and provides training samples for the unit scheduling model module. The multi-dimensional state input module integrates multi-dimensional state variables to form a real-time multi-dimensional state vector and outputs it to the unit scheduling model module. The unit scheduling model module adopts a feedforward neural network structure and outputs qualitative scheduling suggestions L1 to the boundary condition fusion module based on the real-time multidimensional state vector; The boundary condition fusion module integrates the air compressor's operating constraints, transforms the qualitative scheduling suggestion L1 into an operable quantitative scheduling suggestion L2, and outputs it to the operation feedback module. The operation feedback module records staff operations and sends them back to the system, forming a closed-loop data flow to support strategy optimization.

2. The system according to claim 1, characterized in that, The raw operating data includes the air compressor's operating status, main and branch pipe flow rates, and pressure values; the raw operating data is unfiltered and retains all original information from the actual operation process.

3. The system according to claim 1, characterized in that, The evaluation system module evaluates the quality of the initial input data according to preset rules, selects high-quality samples to form a training set, and outputs them to the experience pool module. Specifically, this includes: The evaluation system module comprehensively evaluates the actual effectiveness of historical scheduling operations through a quantitative scoring mechanism. This scoring mechanism is based on two indicators: compliance and volatility within the pressure range. It uses dynamic weight allocation to identify data quality, and the process is as follows: (1) Input data collection: Current status data: including the current system pressure value Real-time flow of main pipe and branch pipes ; Historical moment status data: Pressure value at the previous moment The flow rates of the main pipe and branch pipes at the previous moment and the scheduling scheme executed in the previous time step. ; (2) Interval compliance scoring mechanism: This mechanism measures whether the current pressure value is stable within the target pressure range. The scoring rules are as follows: Assess current stress levels Whether the pressure is within the target pressure range, and deduct points linearly based on the degree of deviation, according to the following rules: ； (3) Volatility scoring mechanism: This mechanism combines pressure and flow rate changes and scores them based on a pre-defined piecewise function; different flow rate change ranges correspond to different pressure fluctuation tolerance thresholds and deduction functions, as detailed below: ； (4) The total score is obtained by adding the above two parts: ； If the total score Then determine the scheduling scheme. If effective, it will be stored in the experience pool for subsequent training of the unit scheduling model and optimization of scheduling strategies.

4. The system according to claim 1, characterized in that, The multidimensional state input module also includes a mechanism flow calculation module, which calculates the mechanism flow through a mechanism fitting curve and combines it with real-time pressure data to form an estimated mechanism flow value that reflects the actual operating state of the system. The estimated flow rate based on the mechanism is then output to the multi-dimensional state input module as part of the input for training the unit scheduling model.

5. The system according to claim 4, characterized in that, The multi-dimensional state input module integrates the mechanism flow estimate, the actual flow of the main pipe and branch pipes, the predicted flow of the main pipe, and the current pressure value to form a real-time multi-dimensional state vector, which is then output to the unit scheduling model module for training.

6. The system according to claim 5, characterized in that, The unit scheduling model module adopts a feedforward neural network structure and outputs qualitative scheduling suggestions L1 based on the real-time multi-dimensional state vector to the boundary condition fusion module, specifically including: A deep learning model based on a feedforward neural network structure is used for air compressor scheduling prediction. The network structure includes one input layer, three hidden layers, and two independent output layers, which together form a multi-task learning framework. (1) Input layer design The network input is a 1×4 dimensional feature vector containing the following four key features: Flow: the current measured flow value; FutureFlow: the predicted flow value in the future period; Pressure: the current pressure value of the system; Mechanism_Flow: the flow value calculated by a professional mechanism model. (2) Hidden layer configuration The network structure consists of three fully connected hidden layers using the ReLU activation function. This network is a multi-layer feedforward neural network that processes data progressively, starting from the input layer: the input layer fc1 receives a 4-dimensional feature vector, which is expanded to 1024 dimensions by the first fully connected layer fc2, and a ReLU activation function is applied to introduce non-linearity; subsequently, the second fully connected layer fc3 compresses the dimension to 512 dimensions, and the third fully connected layer fc4 further compresses it to 256 dimensions. Both fc3 and fc4 layers retain ReLU activation to ensure the model's expressive power; finally, the network structure consists of two independent output layers. (3) Output layer design The network structure employs two independent output layers to predict the load quantities of the two types of air compressors, respectively: The num_big output layer of the type I air compressor: weight dimension: 9×128; bias dimension: 9; outputs a 1×9 vector, representing the probability distribution of the number of large air compressors loaded from 0 to 8. The num_small output layer for the Type II air compressor has the following characteristics: weight dimension: 2×128; bias dimension: 2; and outputs a 1×2 vector representing the probability distribution of the number of small air compressors loaded from 0 to 1.

7. The system according to claim 6, characterized in that, The network hyperparameter settings for the unit scheduling model, including the learning rate settings, are as follows: The learning rate is a core hyperparameter that controls the step size of parameter updates in the training of deep learning models with feedforward neural network structures, and it directly affects the convergence of the model. The Adam optimizer is used, and 0.0004 is selected as the initial learning rate for model training.

8. The system according to claim 7, characterized in that, The network hyperparameter settings also include batch size settings, as follows: Batch size directly affects the training effect and computational efficiency of the model. Based on multiple rounds of experimental comparisons on a medium dataset, a batch size of 256 was determined. This choice fully utilizes the parallel computing power of the GPU to improve training efficiency, maintains sufficient gradient stability to avoid drastic fluctuations during training, and achieves the best balance between memory constraints and training stability. The batch size of the network hyperparameter is set to 256, which, together with the learning rate of 0.0004, achieves gradient stability. When the batch size increases by a factor of k, the gradient variance decreases by a factor of k, so the learning rate needs to be increased simultaneously. This is to maintain the same signal-to-noise ratio during training.

9. The system according to claim 8, characterized in that, During the training of the unit scheduling model, the cross-entropy loss function is used as the optimization objective of the model. Loss values ​​are calculated separately for the prediction tasks of small and large air compressors. Considering the differences between the two types of air compressors in the system, a weighted summation loss function is used, as shown in the following expression: Total loss function value = × Loss value of small air compressor + Loss value of large air compressor By adjusting the weight parameters This can balance the impact of the two types of prediction tasks on the overall optimization process, and set... .

10. The system according to claim 6, characterized in that, The boundary condition fusion module integrates the operating constraints of the air compressors at the air compressor station, transforms the qualitative scheduling suggestion L1 into an operable quantitative scheduling suggestion L2, and outputs the quantitative scheduling suggestion L2 to the operation feedback module; wherein... L1 is the suggested total number of type I and type II air compressors, and L2 is the specific unit and its operation in type I and type II air compressors. The operational constraints include the daily maximum number of start-stop cycles, monthly operational balance requirements, and start-stop interval limits.

Images