A shipyard gas load prediction method and system based on a hybrid network model

Abstract

The application provides a shipyard gas load prediction method and system based on a hybrid network model, and relates to the technical field of intelligent prediction and scheduling of ship energy. The method comprises: collecting field time series data; based on the time series data, analyzing and preprocessing the data characteristics of the shipyard; based on the preprocessed time series data, constructing a hybrid network model GP-ConvFormer with global position bias, modeling long and short term dependencies through a global position bias mechanism; and automatically tuning parameters using a Bayesian optimization framework to achieve high-precision load prediction and dynamic scheduling optimization; combining actual production data of the shipyard, performing example analysis and comparison, comparing the prediction results of the five models, and verifying the effectiveness of the hybrid network model for shipyard gas load prediction; based on the prediction results of the model, industrial scheduling and energy saving optimization are performed. The application can promote fine energy management and energy saving and consumption reduction, and help the intelligent transformation of green shipyards.

Description

Technical Field

[0001] This invention relates to the field of intelligent prediction and scheduling technology for ship energy, specifically to a shipyard gas load prediction method and system based on a hybrid network model. Background Technology

[0002] As global manufacturing transforms towards green and intelligent practices, the shipbuilding industry, being a high-energy-consuming sector, has made refined energy management a crucial link in achieving carbon emission reduction targets and sustainable development. Power gases, essential energy sources for core processes in shipbuilding such as cutting, welding, sandblasting, painting, and hoisting, directly impact shipyard production schedules, product quality, and energy costs due to their supply stability and utilization efficiency. The International Maritime Organization (IMO), in its "Green Shipyard Assessment Guidelines," explicitly states that building sophisticated energy forecasting and management models is a core technological support for the green transformation of the shipbuilding industry.

[0003] However, the compressed air load in shipyards is influenced by multiple factors, including production planning, equipment start-up and shutdown, process switching, shift collaboration, and environmental factors, exhibiting complex characteristics such as strong nonlinearity, high-frequency fluctuations, multimodal transitions, and significant periodicity and phased patterns. Currently, shipyards still largely rely on manual experience for decision-making in predicting power gas flow demand and scheduling air compressors, leading to frequent fluctuations in gas supply pressure, high energy consumption, and low equipment operating efficiency, which seriously restricts the improvement of shipyards' refined energy management level.

[0004] To address these issues, the industry has explored various load forecasting methods. Traditional forecasting methods, such as the Seasonally Differential Autoregressive Moving Average (SARIMA) model, capture the stable trends and seasonal characteristics of time-series data through linear modeling. However, these methods struggle to adapt to the strong nonlinearity and sudden fluctuations in shipyard loads, resulting in limited forecast accuracy. Machine learning-based single models, such as Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and Transformer models, have also been applied. LSTM alleviates the gradient vanishing problem in long-sequence data through gating mechanisms and excels at capturing temporal dependencies, but its responsiveness to local transient abrupt changes is insufficient. CNN, leveraging the local receptive field advantage of convolutional kernels, can effectively extract local detailed features from the data, but it struggles to model long-term, cross-period global dependencies. Transformer, based on self-attention mechanisms, achieves global feature association modeling, but its sensitivity to local abrupt changes is weak, and it is prone to overfitting in small-sample scenarios, resulting in high computational complexity.

[0005] Furthermore, the technological characteristics, production rhythms, and gas demand of different production areas in a shipyard (such as flat workshops, painting and sandblasting areas, and dock platforms) vary significantly, leading to multi-scale feature mixing in load data. Simultaneously, the actual monitoring data inevitably contains data quality issues such as noise interference, missing values, and outliers, further increasing the difficulty of load forecasting. Existing forecasting models do not fully consider the impact of the global locational correlations of shipyard production (such as regional operating conditions, shift schedules, and time-of-day distribution) on load changes, and lack targeted feature enhancement mechanisms, resulting in forecast accuracy and robustness in complex shipyard scenarios that fail to meet actual engineering needs.

[0006] Therefore, there is an urgent need for a high-precision load prediction model that can simultaneously and accurately capture the local instantaneous characteristics and long-term dependencies of load data, and effectively adapt to the global location characteristics and complex data quality issues of shipyard production. This is of great significance for improving the stability of power gas supply in shipyards, reducing energy consumption, and realizing intelligent scheduling. It is also a key issue that needs to be addressed in the current energy management technology of the shipbuilding industry. Summary of the Invention

[0007] To address the over-reliance on human experience in shipyard power gas energy flow demand and air compressor scheduling strategies, the difficulty of traditional linear models (such as SARIMA) to adapt to the strong nonlinearity, high-frequency fluctuations, and multimodal transitions of shipyard gas loads, and the limitations of single machine learning / deep learning models (such as CNN, LSTM, and Transformer) in capturing both local instantaneous features and long-term dependencies, or inadequate consideration of the global location correlations in shipyard production (such as regional operating conditions and shift schedules) and quality issues like data noise and missing values, resulting in insufficient prediction accuracy and robustness, and failing to meet the practical needs of refined energy management and air compressor scheduling in shipyards, this invention proposes a ConvFormer shipyard gas load prediction model method and system that integrates global location bias. The research systematically unfolds from five aspects: shipyard production energy consumption condition classification, data characteristic analysis, data preprocessing, model construction and optimization, and case study analysis. This method exhibits higher robustness in dealing with non-stationarity, multimodal characteristics, and sudden changes in operating conditions, providing technical support for shipbuilding enterprises to achieve intelligent energy system management and energy conservation.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a shipyard gas load prediction method based on a hybrid network model, the method comprising: Conduct on-site time-series data acquisition; Based on the aforementioned time-series data, shipyard data characteristic analysis and preprocessing are performed; Based on preprocessed time-series data, a CNN-Transformer hybrid network model GP-ConvFormer with global position bias is constructed. The global position bias mechanism models long-term and short-term dependencies. The Bayesian optimization framework is used to automatically tune parameters to achieve high-precision load forecasting and dynamic scheduling optimization. By combining actual production data from shipyards, we conducted case studies and comparisons, comparing the prediction results of GP-ConvForme with those of SARIMA, LSTM, CNN, and Transformer models to verify the effectiveness of the GP-ConvForme hybrid network model in predicting shipyard gas loads. Industrial scheduling and energy-saving optimization are performed based on the prediction results of the GP-ConvFormer hybrid network model.

[0009] On the other hand, the present invention also provides a shipyard gas load prediction system based on a hybrid network model. The system includes a memory for storing computer program instructions and a processor for executing the program instructions. When the computer program instructions are executed by the processor, the system is triggered to execute the above-mentioned shipyard gas load prediction method based on a hybrid network model.

[0010] Compared with the prior art, the beneficial effects of the present invention are: 1. Prediction accuracy has been significantly improved. By integrating the advantages of CNN local feature extraction and Transformer long-range dependency modeling, and superimposing the Global Position Bias (GPB) mechanism, this method accurately captures the periodic, phased, and abrupt characteristics of shipyard gas load. The MAE is reduced to 920.5, which is 71.5% lower than traditional SARIMA, 44.4% lower than LSTM, and 35.2% lower than Transformer. The R² reaches 0.98, approaching the theoretical limit, and can explain 98% of the flow variation, solving the systematic bias problem of existing models during operating condition switching periods.

[0011] 2. Stronger adaptability to complex scenarios To address the differentiated gas consumption characteristics of various production areas in shipyards (planar workshops, painting and sandblasting, etc.), a three-dimensional location coding system based on "factory-workgroup-time" is used to adapt to multi-time-varying and multi-scale load characteristics, overcoming the shortcomings of single models in capturing local abrupt changes or long-term trends. Furthermore, through preprocessing processes such as KNN interpolation and adaptive median filtering, common problems in industrial scenarios, such as data noise, missing values, and outliers, are effectively addressed, demonstrating superior robustness compared to traditional models that rely on ideal data.

[0012] 3. Balancing practicality and efficiency in engineering A Bayesian-optimized hyperparameter automatic optimization framework, coupled with an early-stop strategy, enhances the model's generalization ability under different shipyard operating conditions, eliminating the need for repeated manual parameter tuning. The average prediction time per step is only 480ms, lower than the Transformer's 600ms, meeting the real-time scheduling requirements of industry while balancing high accuracy and computational efficiency.

[0013] 4. Supports energy-saving scheduling and refined management High-precision forecasting results provide a reliable basis for the dynamic scheduling of air compressors, enabling early prediction of load peaks and troughs, and reasonable adjustment of the number and power of equipment starting and stopping, avoiding redundant operation and reducing pipeline pressure fluctuations. It also helps optimize the operating logic of air tanks and flow control valves, reducing equipment start-up and shutdown frequency and no-load energy consumption. Furthermore, through unplanned disturbance early warning, it avoids energy waste under abnormal operating conditions, providing technical support for shipyards' energy conservation, emission reduction, and green transformation.

[0014] 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

[0015] Figure 1 This is a flowchart of the GP-Convformer modeling method according to the present invention; Figure 2 This is a time-series characteristic decomposition diagram (one week) of the gas load curve of a typical production area according to the present invention. Figure 3 This is a curve diagram of the original data collected in the shipyard dock area according to the present invention; Figure 4 This is a data curve diagram of the shipyard dock area according to the present invention; Figure 5 This is a schematic diagram of the GP-ConvFormer hybrid network model according to the present invention; Figure 6 This is a graph showing the change in the training loss function of the GP-Convformer model according to the present invention; Figure 7 This is a comparison chart of 6-hour flow prediction using five models according to the present invention; Figure 8 This is a comparison chart of SARIMA model predictions; Figure 9 This is a comparison chart of predictions from the LSTM model; Figure 10 This is a comparison chart of CNN model predictions; Figure 11 This is a comparison chart of predictions from the Transformer model; Figure 12 This is a comparison of predictions made using the GP-Convformer model according to the present invention. Detailed Implementation

[0016] 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.

[0017] 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”.

[0018] 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.

[0019] 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.

[0020] 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.

[0021] 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.

[0022] To address the aforementioned technical problems of insufficient accuracy and robustness in shipyard gas load forecasting, which hinders its ability to meet the practical needs of refined energy management and air compressor scheduling in shipyards, this invention proposes a shipyard gas load forecasting model based on a GP-ConvFormer hybrid network model incorporating global position bias. This model innovatively combines the advantages of Convolutional Neural Networks (CNNs) in local feature extraction with the capabilities of Transformers in long-range dependency modeling, and introduces a learnable global position bias (GPB) mechanism to enhance the ability to capture the periodic and phased characteristics of shipyard gas load data.

[0023] To fully consider the energy consumption characteristics of shipyard production and the gas load data characteristics of each production area, this invention proposes a shipyard gas load prediction model method that integrates a global location bias GP-ConvFormer hybrid network model, such as... Figure 1 As shown.

[0024] S1. Collect on-site time-series data; To comprehensively acquire information on the supply and consumption of power gases in the shipyard, a hardware architecture centered on PLC controllers and industrial IoT gateways is used. Through Profibus-DP and Modbus RTU industrial buses, real-time monitoring is conducted on air compressors and related equipment in all core production areas, including the flat workshop, painting and sandblasting areas, and the dock platform. At a sampling frequency of 1Hz, precise data is collected down to the second level, including timestamps (data_time), instantaneous flow values ​​(data_flow), unique device identifiers (data_devicecode), and key operating parameters such as air compressor exhaust pressure, operating current, start / stop status, and electric regulating valve opening. The collected time-series data is uploaded to the backend management platform in real-time via the MQTT protocol. It is then structured and stored according to the "device code-timestamp" dimension to form a standardized raw time-series database. Simultaneously, it covers scenarios with multiple devices operating in parallel, establishing multiple independent data acquisition channels to achieve accurate traceability and differentiation of gas consumption data from different areas and devices. This provides comprehensive, accurate, and effective data support for subsequent data preprocessing and model training.

[0025] S2. Based on the time series data, perform shipyard data characteristic analysis and preprocessing; Based on the shipbuilding process, the shipyard is divided into different energy-consuming areas and operating conditions. On this basis, time-series decomposition and load data characteristic analysis are performed on gas load data from typical shipyard production areas to identify data quality issues in actual shipyard gas load scenarios. Targeted data preprocessing methods are proposed: KNN interpolation and adaptive filtering are used to remove noise and missing values ​​from the original time-series data; and absolute position encoding and multi-granularity periodic feature engineering are combined to enhance the model's ability to capture the time-series characteristics of gas consumption in shipyards.

[0026] Specifically, based on the shipbuilding process, the shipyard is divided into different energy-consuming areas for production. Time-series decomposition is performed on the gas load data of typical production areas in the shipyard (see...). Figure 2 The analysis reveals that the shipyard's overall power gas system exhibits a clear periodicity, regularity, and phased trend, but the load characteristics vary significantly between different areas. The planar workshops, primarily engaged in cutting and welding, have stable loads with daily load fluctuations of less than 5%, exhibiting a standardized and continuous rhythm. The painting and sandblasting area has a prominent high-pressure demand, with an average pressure of approximately 7.5 bar, showing a clear nighttime off-peak characteristic, with peak loads occurring mostly after 8:00 PM. The dock platform, due to overlapping processes and strong operational randomness, exhibits irregular and sudden fluctuations in gas consumption. The painting, 300T platform, and curved surface workshops show strong operational periodicity, but gas demand varies significantly at different stages. Therefore, while the overall shipyard load exhibits a periodic pattern, local areas are significantly affected by processes and production organization methods, demonstrating multi-temporally variable and multi-scale gas consumption characteristics.

[0027] In addition, shipbuilding involves various stages in production areas such as painting, flat surface workshops, curved surface workshops, 300T platforms, sandblasting, and dock platforms. Due to the different technological characteristics, equipment configurations, and operating modes of each stage, there are certain differences in gas consumption and load characteristics in different production areas of the shipyard. The main energy consumption conditions of the shipyard are shown in Table 1.

[0028] Table 1 Operating conditions of the main production areas of the shipyard However, in actual production processes, the acquisition and transmission of gas load data are susceptible to various external interference factors. On the one hand, planned fluctuations stem from routine changes such as production scheduling, shift work schedules, and work section handovers, resulting in different load patterns on different dates within the same time period. To effectively capture these regular but unpredictable fluctuation trends and incorporate them into model training, this invention introduces a Global Positional Bias (GPB) mechanism to encode the location semantics of the "factory-shift-time" dimension into data features, enabling deep learning models to identify such systematic fluctuations caused by scheduling or work schedules. On the other hand, unplanned disturbances (such as equipment failure, sensor drift, network interruption, and pipeline pressure fluctuations) introduce random anomalies, causing load data to exhibit typical sensor noise characteristics. These manifest as short-duration impulse noise (usually lasting less than 1 second), persistent offsets (equipment malfunctions leading to overestimation or underestimation of measurements), zero values, or invalid values ​​(communication anomalies or data loss). These anomalies are mixed with actual operating conditions, making it difficult for traditional detection methods based on fixed thresholds or simple moving averages to accurately distinguish them.

[0029] Furthermore, the flow characteristics differ significantly across different production stages. For example, during normal processing, equipment start-up / shutdown, and maintenance, the load signals exhibit significant differences in mean, variance, and frequency domain characteristics. Particularly during operational transitions, data sequences often exhibit non-stationary characteristics, manifesting as abrupt changes, jumps, or short-term drifts. Without proper preprocessing, these characteristics directly impact the model's time-series pattern learning and feature extraction accuracy, leading to systematic biases in prediction results. Therefore, scientific data preprocessing before entering the modeling stage is crucial for ensuring model performance. This invention, based on a thorough consideration of the differences in shipyard production areas and data noise characteristics, designs a multi-layered data preprocessing workflow, including missing value repair, anomaly detection and smoothing filtering, feature normalization, and periodic signal extraction. By introducing methods such as KNN interpolation, adaptive median filtering, and adaptive distribution elimination, dynamic repair and stabilization of sudden anomalies are achieved. At the same time, by combining the "production area location" and "shift-time period location" in the GPB mechanism, these location information are encoded into the time series data. In the subsequent "GP-ConvFormer" model training stage, the model can not only learn the "load value sequence itself change", but also use the auxiliary feature of "the position of the data point in the global production rhythm" to better fit the fluctuation pattern brought about by production scheduling.

[0030] Next, the collected raw time-series data is preprocessed. This raw time-series data contains three core fields: a timestamp accurate to the second (data_time), instantaneous flow rate (data_flow), and a unique device identifier (data_devicecode). During the data loading stage, the pandas library's `read_csv` function is used for reading, and Boolean indexing is used to efficiently remove invalid records containing NaN values, ensuring that subsequent processing is based on a complete data matrix. Addressing the common need for parallel monitoring of multiple devices in industrial scenarios, the system establishes seven independent data processing channels based on device coding characteristics. Automatic data splitting is achieved by traversing the device coding list, with each channel corresponding to a specific device's data set. To address various issues with the raw time-series data, this invention employs methods such as KNN interpolation, adaptive median filtering, and absolute positional encoding to process the raw time-series data into smoothed data with positional encoding suitable for subsequent model training.

[0031] Specifically, 1) the KNN interpolation method is used: The data alignment and normalization module employs a K-nearest neighbor (KNN) based interpolation method (data_filling function) to address the non-uniform sampling problem. This algorithm sets a standard time length of 3 days (3 × 24 × 60 × 60 seconds). When the input data length exceeds this standard value, redundant sampling points are deleted at equal time intervals. When data is insufficient, linear interpolation is used to generate the arithmetic mean of two adjacent points as a new data point. This approach ensures time alignment while preserving the trend characteristics of the original time-series data to the greatest extent possible.

[0032] 2) Adaptive median filtering algorithm is used: This invention employs an improved adaptive median filtering algorithm to handle outliers. A 300-second sliding window is used for local statistical analysis. First, the mean of the data within the window is calculated as a benchmark. Data points are then sorted by calculating the variance of each data point relative to the mean, and the top 40% of data with the largest variance (p=0.4) are dynamically removed. The remaining data is filled in using KNN interpolation. The algorithm innovatively introduces a skewness detection mechanism. When the data distribution is uneven around the mean, a secondary filtering is triggered, retaining only data points above the mean. Compared to traditional median filtering, this algorithm implements two levels of adaptive processing: the first level accurately identifies outliers through variance analysis, avoiding the loss of effective data caused by region smoothing; the second level intelligently judges feature edges based on distribution morphology, automatically switching to a conservative mode when an edge is detected, removing only obvious outliers. This design balances noise removal and feature protection, minimizing signal distortion while ensuring effective denoising.

[0033] 3) Absolute position encoding: Shipyard gas consumption is periodic; for example, gas consumption stops when operations cease at night. To capture this pattern, this invention employs a location-based encoding method. Time features are encoded using three different granularities of trigonometric functions: minute-level period (sin(2πt / 60)), hour-level period (cos(2πt / 1440)), and composite periodic period (w%7 / 7+t%10080 / 10080). This encoding method provides better periodic continuity compared to the original timestamp, which is beneficial for the model to capture the temporal patterns of equipment operation. Flow rates, after Z-score standardization, are combined with binarized status flags to form the input features. The status threshold is dynamically set according to the equipment type (thres_list parameter), ranging from 10 to 10000.

[0034] By adding time and state features of the same length to the original 288-bit data, it is expanded into a 288×4 data matrix. The processed data matrix effectively eliminates the influence of on-site noise, preserves key temporal features, and provides a regularized input tensor for subsequent deep learning models such as the Transformer. The final input data dimension for the deep learning model is [N, 4, 864], where N is the number of samples, the four channels correspond to flow rate, state flag, sine encoding, and cosine encoding, respectively, and 864 represents the windowing interval of 5 minutes every three days. Each data point has an input length of 24×12×2=576 and an output length of 24×12=288. The processed matrix form is as follows.

[0035] X= Where, x i This represents the shipyard's gas load value; 0 / 1 indicates the current operating status of the equipment, with 0 indicating the equipment is off and 1 indicating the equipment is on; hour i This is the hour value for the current moment.

[0036] Comparison of data before and after processing, for example Figure 3 and 4 As shown, high-frequency noise in the original time-series data has been removed, and the curve has become relatively smooth. Obvious outliers caused by data acquisition issues have also been corrected.

[0037] S3. Construct the CNN-Transformer hybrid network model GP-ConvFormer with global position bias, model long-short-term dependencies through the global position bias mechanism, and use the Bayesian optimization framework to automatically tune parameters to achieve high-precision load forecasting and dynamic scheduling optimization. Specifically, ConvFormer is a hybrid model that combines the advantages of CNN and Transformer, specifically optimized for time series prediction tasks. Its core architecture deploys a CNN module and a Transformer encoder in parallel, simultaneously capturing local details (such as short-term fluctuations and abrupt changes) and global dependencies (such as long-term trends and cross-variable associations) in sequence data. The CNN module uses six convolutional layers to compress the sequence length and extract local patterns, while the Transformer encoder establishes long-range time-step associations through multi-layer self-attention mechanisms. Finally, feature concatenation and fully connected layers achieve multi-scale feature fusion, significantly improving the model's performance in time series tasks such as load prediction and equipment failure early warning.

[0038] The innovation of this ConvFormer model lies in: 1) enhancing sensitivity to local temporal abrupt changes (such as sensor noise and transient anomalies) through convolutional operations, compensating for the shortcomings of pure Transformer in fine-grained feature extraction; 2) retaining the advantages of Transformer in handling long-range dependencies, solving the problem that traditional RNNs struggle to model annual / quarterly cycles due to gradient vanishing; 3) the feature fusion mechanism automatically balances the contribution weights of local and global information, exhibiting stronger robustness in complex industrial scenarios. Experiments demonstrate that compared to single-architecture models, ConvFormer significantly improves both prediction accuracy and computational efficiency. The Global Positional Bias (GPB) mechanism introduced in this invention further enhances the model's ability to model the systematic fluctuations in shipyard production rhythm, forming the GP-ConvFormer model. GPB outputs positional bias weights with a length of 288. Combine the predictions from CNNs with those from Transformers.

[0039] In the formula, For the final prediction result, and These are the prediction results from CNN and Transformer, respectively.

[0040] The next step is to optimize the hyperparameters of the GP-ConvFormer hybrid network model. Hyperparameter optimization is a crucial step in ensuring model performance during training. This invention employs a hierarchical optimization strategy, dividing hyperparameters into three interrelated categories: architecture parameters (including kernel size, number of attention heads, and hidden layer dimensions), training parameters (covering learning rate, batch size, and dropout rate), and Gaussian process parameters (involving kernel function type, length scale, and output scale). This hierarchical approach not only improves search efficiency but also ensures collaborative optimization among parameters.

[0041] This invention constructs an automatic hyperparameter tuning framework based on Bayesian optimization, establishing a probabilistic surrogate model between the objective function and hyperparameters through a tree-structured Parzen estimator (TPE). Specifically, a search space comprising 50 trials is defined, where the learning rate adopts a log-uniform distribution (1e-5 to 1e-2), the number of attention heads is discretely selected from {4, 8, 16}, and the convolutional kernel size is optimized within the range {3, 5, 7}. An early stopping strategy is used for each trial, terminating training when the validation set loss fails to improve within 10 consecutive epochs, significantly improving search efficiency.

[0042] like Figure 6 As shown, the GP-ConvFormer hybrid network model exhibits excellent training convergence characteristics. Both the training loss and validation loss show a strictly monotonically decreasing trend, with the training loss continuously decreasing from the initial value of 0.85 to 0.05 (a reduction of 94.1%), while the validation loss maintains a reasonable gap of 15-20%. This smooth, non-oscillating exponential decay curve indicates that: 1) the adaptive learning rate scheduling strategy adopted by the model effectively avoids gradient abrupt changes; 2) the synergistic design of graph convolution and self-attention mechanisms ensures the stability of parameter updates. This demonstrates that the spatiotemporal feature fusion mechanism of the GP-ConvFormer hybrid network model significantly improves the generalization ability of gas flow load prediction.

[0043] S4. Combine actual production data from the shipyard to conduct case studies and comparisons. Compare the prediction results of GP-ConvFormer with those of SARIMA, LSTM, CNN, and Transformer models to verify the effectiveness of the shipyard gas load prediction method using the GP-ConvForme hybrid network model provided in this invention. Specifically, 1) SARIMA model The Seasonal Autoregressive Integrated Moving Average (SARIMA) model combines autoregressive (AR), differencing (I), and moving average (MA) methods, and incorporates seasonality (S) to effectively predict time series data with clear seasonal patterns and stationarity. The definition is as follows: (1) in, It is a non-seasonal AR polynomial. It is a seasonal AR polynomial. It is a non-seasonal MA polynomial. It is a seasonal MA polynomial, and B is the delay operator. It is time series data. It is white noise, and S is the seasonal period.

[0044] The parameters of SARIMA are typically expressed as (p,d,q)(P,D,Q)(m), where p, d, and q are the autoregressive, differencing, and moving average orders of the non-seasonal component, respectively; P, D, and Q are the autoregressive, differencing, and moving average orders of the seasonal component, respectively; and m represents the length of the seasonal cycle. In this experiment, the parameters were chosen as (0,0,2)(2,0,0,48).

[0045] 2) LSTM model Long Short-Term Memory (LSTM) is a special type of recurrent neural network (RNN) primarily used to address the vanishing gradient problem in traditional RNNs when processing long sequences of data. Its core innovation lies in the introduction of gating mechanisms (including forget gates, input gates, and output gates), which effectively capture long-term dependencies in time series data by selectively remembering and forgetting information. This structure makes LSTMs superior in tasks such as time series prediction. Compared to traditional RNNs, LSTMs can handle dependencies over 100 time steps, and compared to newer architectures like the Transformer, they offer higher computational efficiency on smaller datasets. Variants of LSTM, such as Bidirectional LSTM (BiLSTM) and LSTMs incorporating attention mechanisms, further enhance model performance, making it one of the important foundational architectures for sequence modeling.

[0046] 3) CNN model A Convolutional Neural Network (CNN) consists of convolutional layers, pooling layers, activation functions, and other components. These components work together to perform sliding window-based operations on the input data using convolutional kernels, thereby analyzing local features of different sizes and locations and combining them into a higher-level feature representation.

[0047] This mechanism gives CNNs a natural advantage in time series data with local correlations, such as vibration signals, which require automated feature extraction. However, the local receptive field of the convolutional kernel limits the capture of long-term patterns (such as seasonal trends), requiring deep network stacking, while deep CNNs are prone to the gradient vanishing problem. The core computation of CNNs is usually implemented through convolution and pooling operations, and their general formulas are shown in equations (2) and (3): (2) (3) in, A convolution kernel of a certain length, Here, F(x) represents the bias term, and F(x) represents the activation function. Poolings(x) represents the output features after processing by the convolutional layer.

[0048] 4) Transformer model The Transformer is a model based on the self-attention mechanism, which establishes interrelationships between elements at different positions in the input sequence, thereby more effectively capturing dependencies between sequence elements and forming more accurate and efficient feature representations. This model mainly consists of an encoder, decoder, self-attention mechanism, multi-head attention mechanism, and feedforward neural network. Its advantage lies in its ability to directly model global dependencies between arbitrary time steps, effectively solving the problem of long-period pattern capture caused by gradient vanishing in traditional RNNs or LSTMs. However, self-attention is weaker than one-dimensional convolution in capturing local abrupt features (such as equipment failure pulse signals), requiring combination with CNNs to enhance local sensitivity. Furthermore, it is prone to overfitting in small sample scenarios and requires a large amount of training data.

[0049] Next, to systematically evaluate the performance of the GP-ConvFormer model, it is compared with the four types of models mentioned above. These four types of models represent different modeling approaches and can comprehensively test the overall performance of GP-ConvFormer in industrial scenarios. This invention takes the air compressor station of a large shipyard in Shanghai as the research object, selecting the operating data of the shipyard's air compressor station from May 1st to 11th, 2025 (11 days in total) as the training set, and from May 12th to 14th, 2025 (3 days in total) as the test set. The data acquisition system is based on a PLC controller and an industrial IoT gateway, acquiring key parameters such as air compressor exhaust pressure, operating current, start / stop status, and electric regulating valve opening in real time through Profibus-DP and Modbus RTU industrial buses, with a sampling frequency of 1Hz. After data preprocessing, it is uploaded to the management platform via the MQTT protocol to form a standardized dataset. Based on actual production scheduling needs, a gas load prediction model is finally constructed with a time step of 5 minutes. We employed a strict variable control approach to ensure that all models were compared under the same data partition (80% training set, 10% validation set, and 10% test set), a unified hyperparameter optimization strategy (based on Bayesian search), and consistent evaluation metrics (MAE, MSE, R²).

[0050] For performance comparisons of various models, this invention employs both quantitative index comparisons and qualitative visualization comparisons: 1) Comparison of quantitative indicators: By comparing the evaluation metrics of the five models (see Table 2), it can be found that the different models exhibit significant differences in the traffic forecasting task. Among them, the SARIMA model has a high MAE of 3223.8 and an MSE of 4.27 × 10⁻⁶. 7This reflects the inherent limitations of traditional linear methods in industrial nonlinear systems. Its R²=0.84 indicates that it can only explain 84% of the flow variation, especially exhibiting systematic biases during operating condition switching. The CNN model achieves an MAE of 2350.0 (a 27.1% reduction compared to SARIMA) and a 33.3% decrease in MSE, validating the ability of convolutional operations to capture instantaneous flow fluctuations in equipment. However, its R²=0.94 is still lower than LSTM, revealing its inadequacy in long-range dependency modeling. LSTM establishes a benchmark for sequence modeling with an MAE of 1656.1, and its gating mechanism further reduces the MSE by 57.5% compared to CNN, exhibiting better temporal memory characteristics, particularly during continuous gas supply phases with t>3h. While the Transformer model further compresses the MAE to 1420.0 (a 14.3% improvement over LSTM), its 600ms single-step time exposes the engineering cost of the O(n²) complexity of the self-attention mechanism. It's worth noting that its R²=0.96, the same as LSTM, suggests that it has reached the bottleneck of traditional architectures in terms of interpretability. GP-ConvFormer, however, achieves state-of-the-art performance with a MAE of 920.5, further reducing the error by 35.2% compared to Transformer. Its MSE is 4.5 × 10⁻⁶. 6 It achieves only 10.5% of the performance of SARIMA, and its R² = 0.98 closely approximates the theoretical limit. The model's average prediction time per step is 480ms, meeting the real-time requirements of industrial scenarios.

[0051] Table 2 Comparison of Evaluation Indicators for Five Models 2) Qualitative visual comparison: like Figure 7 As shown, the load prediction performance of five models over a six-hour period was compared. SARIMA's prediction curve (blue) exhibits significant lag and smoothing effects, particularly noticeable during welding start-up and shutdown (t=0.5h, 2.5h). LSTM (yellow) improves upon the lag issue, but gradually accumulates biases during continuous operation (t>4h). CNN (green) responds quickly to instantaneous changes, but exhibits excessive fluctuations in long-term predictions. Transformer (red) captures periodic features well, but overshoot occurs at abrupt change points. GP-ConvFormer (purple) performs best overall, accurately following the peak changes during welding torch start-up and shutdown while maintaining stable tracking during continuous gas supply. It shows no lag or distortion at any operating condition transition points, and its curve almost perfectly matches the actual value (black), validating the advantages of the hybrid architecture in multi-scale dynamic modeling of shipyard gas loads.

[0052] The detailed prediction comparisons of each model are shown in the following series of figures ( Figures 8-12 This further demonstrates their respective characteristics and limitations.

[0053] like Figure 8 As shown, the SARIMA prediction trajectory reveals the inherent limitations of classical time-series methods: its linear combination characteristics result in a significant phase lag in the response curve and a systematic smoothing bias at points of abrupt changes in operating conditions. This stems from the model's insufficient ability to capture nonlinear dynamics—while differential operations can handle trend terms, they cannot adapt to transient processes with sudden changes in air compressor flow. Furthermore, its fixed-parameter structure makes it difficult to model the complex multi-condition switching patterns of shipyards. This structural deficiency directly manifests as a persistent under-adjustment of predicted values ​​at process transition points (such as welding start-up and shutdown times).

[0054] like Figure 9 As shown, the prediction curves of the LSTM model exhibit significant advantages in sequence modeling: its gating mechanism effectively captures the temporal dependence of the air compressor flow rate, and the LSTM model captures the overall trend of the flow rate well. However, the model still has two typical shortcomings: first, there is a slight delay of about 1-2 sampling points in the response to sudden changes in flow rate, reflecting the insufficient efficiency of the memory unit in extracting transient features; second, asymptotic error accumulation occurs during the continuous air supply phase (t>4h), exposing the inherent long-term memory decay problem of recurrent neural networks. These limitations provide empirical evidence for improvement schemes such as introducing attention mechanisms.

[0055] like Figure 10 As shown, the prediction results of the CNN model exhibit a clear dependence on local features: during the rapid fluctuation phase of the flow, the model can accurately capture short-term abrupt changes (peak-valley error is reduced by about 40% compared to SARIMA), thanks to the local receptive field characteristics of the convolutional kernel. However, its prediction curve gradually deviates from the true value during the long-cycle gas supply phase of t>3h, revealing a fundamental deficiency of lacking temporal memory ability—although convolutional operations can extract instantaneous features, they cannot model dynamic relationships across time steps.

[0056] like Figure 11 As shown, the prediction curves of the Transformer model exhibit good global feature capture capabilities: its self-attention mechanism effectively models long-cycle operating condition dependencies, and the prediction error is reduced by about 15% compared to LSTM during the continuous gas supply phase of t>2h. However, the predicted values ​​show irregular fluctuations during the stable load phase, revealing the lack of modeling of first-principles constraints of industrial systems by the pure attention mechanism, providing a key direction for improvement in subsequent hybrid architecture design.

[0057] The prediction curve of the GP-ConvFormer model is as follows: Figure 12As shown, this model has demonstrated its advantages in modeling operational data for shipyard air compressor stations. The model effectively captures the complex coupling characteristics of multiple air compressors operating in parallel through a Gaussian process component, and its covariance function design is particularly suitable for describing the nonlinear relationship between pressure, flow rate, and power in the air supply system. The convolution module accurately identifies the flow rate mutation patterns caused by typical operating condition changes such as welding torch start-up and shutdown, while the attention mechanism successfully models the correlation patterns of air demand at different times. This hybrid architecture overcomes the key bottleneck of traditional methods in air compressor station optimization: accurately reflecting instantaneous changes in air demand while maintaining continuous tracking of the coordinated operating status of each unit. The high-precision predictions output by the model provide a reliable decision-making basis for subsequent load allocation and start-up / shutdown strategy optimization, especially performing excellently when dealing with intermittent high-load conditions such as hull section welding. Its prediction results show significantly better agreement with actual air compressor station operating data than existing methods, verifying the applicability of this method in specific air compressor station scenarios.

[0058] S5. Industrial scheduling and energy-saving optimization based on prediction results of GP-ConvFormer hybrid network model.

[0059] After the GP-ConvFormer model has demonstrated high-precision predictive performance through core indicators such as MAE, MSE, and R², its prediction results can be deeply applied to the industrial scheduling and energy-saving optimization practices of shipyard air compressor stations. Based on the multi-regional gas load time-series prediction curves output by the GP-ConvFormer model, and combined with the differences in operating conditions and pressure demand characteristics of various production areas (such as the continuous stable load of the flat workshop, the nighttime peak load of painting and sandblasting, and the sudden peak load of the dock platform), a dynamic scheduling strategy is constructed: by predicting and anticipating peak and off-peak periods of load in advance, the number of air compressors started and stopped and their operating power are reasonably adjusted. While meeting the process requirements such as the 7.5-8 bar high-pressure demand of painting and sandblasting and the instantaneous peak load compensation of the dock platform, redundant operation of multiple devices is avoided. In view of the differences in shift collaboration modes and working hours in different areas, the production rhythm patterns captured by the global position bias mechanism are used to achieve the coordinated allocation of cross-regional gas supply resources, reducing pipeline pressure fluctuations and energy waste. In terms of energy conservation and optimization, the scheduling of gas storage tanks and the adjustment logic of diversion valves are optimized based on accurate gas load forecasting results, reducing the frequency of equipment start-up and shutdown and the idle running time, and reducing energy consumption caused by unstable gas supply pressure. At the same time, by combining the characteristics of unplanned disturbances (such as equipment failure and sensor drift) identified by the GP-ConvFormer model, maintenance warnings and emergency scheduling plans are triggered in advance to avoid inefficient energy consumption under abnormal operating conditions. Ultimately, the goal of accurate matching of supply and demand, intelligent and efficient scheduling, and energy conservation and consumption reduction of the shipyard's power gas system is achieved, promoting the transformation of green shipyards to refined energy management.

[0060] The present invention also provides a shipyard gas load prediction system based on a hybrid network model. The system includes a memory for storing computer program instructions and a processor for executing the program instructions. When the computer program instructions are executed by the processor, the system is triggered to execute the above-mentioned shipyard gas load prediction method based on a hybrid network model.

[0061] This invention provides a shipyard gas load prediction method and system based on a hybrid network model. Addressing the complex characteristics of shipyard gas load—strong nonlinearity, high-frequency fluctuations, and multimodal transitions—and the problems of existing prediction models relying on human experience, difficulty in balancing local instantaneous features and long-range dependencies, and insufficient consideration of global location correlations, this invention utilizes on-site time-series data acquisition. Data preprocessing is performed using methods such as KNN interpolation, adaptive median filtering, and absolute position encoding. A GP-ConvFormer hybrid network model is constructed, integrating the advantages of CNN local feature extraction, Transformer long-range dependency modeling capabilities, and the Global Position Bias (GPB) mechanism. A Bayesian optimization framework is employed to achieve automatic hyperparameter optimization. Comparative experiments with SARIMA, LSTM, CNN, and Transformer models demonstrate that the GP-ConvFormer model achieves a MAE of 920.5 and an R² of 0.98, significantly outperforming existing models in prediction accuracy while meeting industrial real-time requirements. This invention can accurately capture the periodic, phased, and abrupt characteristics of gas load in multiple areas of a shipyard, effectively solving quality problems such as data noise and missing values. Based on its prediction results, a dynamic scheduling strategy can be constructed to achieve collaborative allocation of gas supply resources across regions, optimize the operating logic of gas storage tanks and diversion valves, reduce redundant equipment operation and idle energy consumption, and provide early warning of unplanned disturbances. Ultimately, it can improve the stability of power gas supply in shipyards, promote refined energy management and energy conservation, and help green shipyards achieve intelligent transformation.

[0062] 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 gas load prediction method based on a hybrid network model, characterized in that, The method includes: Conduct on-site time-series data acquisition; Based on the aforementioned time-series data, shipyard data characteristic analysis and preprocessing are performed; Based on preprocessed time-series data, a CNN-Transformer hybrid network model GP-ConvFormer with global position bias is constructed. The global position bias mechanism models long-term and short-term dependencies. The Bayesian optimization framework is used to automatically tune parameters to achieve load forecasting and dynamic scheduling optimization. By combining actual production data from shipyards, we conducted case studies and comparisons, comparing the prediction results of GP-ConvForme with those of SARIMA, LSTM, CNN, and Transformer models to verify the effectiveness of the GP-ConvForme hybrid network model in predicting shipyard gas loads. Industrial scheduling and energy-saving optimization are performed based on the prediction results of the GP-ConvFormer hybrid network model.

2. The method according to claim 1, characterized in that, The analysis of shipyard data characteristics based on the time-series data specifically includes: Based on the shipbuilding process, the shipyard was divided into different energy-consuming areas. By performing time-series decomposition of gas load data from typical production areas, significant differences in load characteristics were found between different areas: the planar workshop, mainly engaged in cutting and welding, had a stable load with daily load fluctuations of less than 5%, exhibiting a standardized and continuous rhythm; the painting and sandblasting area had a prominent high-pressure demand, with an average pressure of 7.5 bar, showing obvious nighttime off-peak characteristics, with peak loads mostly occurring after 8:00 PM; the dock platform, due to the overlapping of processes and the high randomness of operations, exhibited irregular and sudden fluctuations in gas consumption; while the painting, 300T platform, and curved surface workshops had strong operational cycles, but the gas demand varied significantly at different stages.

3. The method according to claim 1, characterized in that, The shipyard data preprocessing based on the time-series data specifically includes: For the collected raw time-series data, KNN interpolation, adaptive median filtering, and absolute position encoding methods are used to process the raw time-series data into smooth data with position encoding suitable for subsequent model training. 1) Using the KNN interpolation method: The data alignment and standardization module uses a K-nearest neighbor (KNN) interpolation method to handle non-uniform sampling problems. This method is implemented through the data_filling function. The data_filling function sets the standard time length to 3 days. When the length of the input data exceeds the standard value, redundant sampling points are deleted at equal time intervals. When the data is insufficient, the arithmetic mean of two adjacent points is generated by linear interpolation as a new data point. 2) Adaptive median filtering algorithm is used: A 300-second sliding window is set up for local statistical analysis. First, the mean of the data within the window is calculated as a benchmark. The data points are then sorted by calculating the variance of each data point relative to the mean. The top 40% of data with the largest variance are dynamically removed, and the remaining data is filled in using KNN interpolation. 3) Absolute position encoding: The time features are encoded using three different granularities of trigonometric functions: the sine function sin(2πt / 60) with a minute-level period, the cosine function cos(2πt / 1440) with an hour-level period, and the normalized calculation formula w%7 / 7+t%10080 / 10080 for the composite period. By adding time and state features of the same length to the original 288-bit data, it is expanded into a 288×4 data matrix. The processed data matrix eliminates the influence of on-site noise, retains key time-series features, and the final output data dimension is [N, 4, 864], where N is the number of samples, the four channels correspond to flow rate, status flag, sin encoding, and cos encoding, respectively, and 864 represents the windowing interval of 5 minutes every three days. The input of each data point is 24×12×2=576 bits, and the output is 24×12=288 bits. The processed matrix form is as follows: X= ； ； Where, x i This represents the shipyard's gas load value; 0 / 1 indicates the current operating status of the equipment, with 0 indicating the equipment is off and 1 indicating the equipment is on; hour i This is the hour value for the current moment.

4. The method according to claim 1, characterized in that, The construction of the CNN-Transformer hybrid network model GP-ConvFormer with global position bias, which models long-short-term dependencies through a global position bias mechanism, specifically includes: ConvFormer is a hybrid model that combines CNN and Transformer, optimized for time series prediction tasks. The core architecture of this hybrid model involves parallel deployment of CNN modules and a Transformer encoder, simultaneously capturing local details and global dependencies in the sequence data. The CNN module uses six convolutional layers to compress the sequence length and extract local patterns, while the Transformer encoder establishes long-range time step relationships through multi-layer self-attention mechanisms. Finally, multi-scale feature fusion is achieved through feature concatenation and fully connected layers. Furthermore, a Global Position Bias (GPB) mechanism is introduced to further enhance the model's ability to model the systematic fluctuations in shipyard production rhythms, forming the GP-ConvFormer model. GPB outputs position bias weights with a length of 288. The predictions from the CNN and the Transformer are combined using the following expression: ； In the formula, For the final prediction result, and These are the prediction results from CNN and Transformer, respectively.

5. The method according to claim 4, characterized in that, The automatic parameter tuning using the Bayesian optimization framework to achieve high-precision load forecasting and dynamic scheduling optimization specifically includes: An automatic hyperparameter tuning framework was constructed based on Bayesian optimization. A probabilistic surrogate model between the objective function and hyperparameters was established through a tree-structured Parzen estimator (TPE). In the specific implementation, a search space containing 50 trials was defined, where the learning rate adopts a log-uniform distribution with a value range of 1e-5 to 1e-2, the number of attention heads is discretely selected from {4, 8, 16}, and the convolutional kernel size is optimized within the range of {3, 5, 7}. Furthermore, an early stopping strategy is used for each trial, terminating training when the validation set loss does not improve within 10 consecutive epochs, thereby improving search efficiency.

6. The method according to claim 5, characterized in that, By combining actual production data from shipyards, a case study analysis and comparison were conducted. The prediction results of GP-ConvFormer were compared with those of SARIMA, LSTM, CNN, and Transformer models to verify the effectiveness of the GP-ConvFormer hybrid network model in predicting shipyard gas loads. Specifically, this included: To systematically evaluate the performance of the GP-ConvFormer model, it was compared with four other models: SARIMA, LSTM, CNN, and Transformer. A period of shipyard air compressor station operation data was selected as the training set, and another period as the test set. After preprocessing, the data was uploaded to the management platform via the MQTT protocol to form a standardized dataset. A gas load prediction model was constructed with a 5-minute time step. Strict variable control was employed to ensure that all models were compared under the same data partitioning, a unified hyperparameter optimization strategy, and consistent evaluation metrics including mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²). The data was divided into a training set (80%), a validation set (10%), and a test set (10%). For the performance comparison of various models, quantitative indicators and qualitative visualization were used. Through comparative analysis, the GP-ConvForme hybrid network model showed the best performance among the five types of models.

7. The method according to claim 6, characterized in that, The method of using quantitative indicators for comparison and qualitative visualization for comparison specifically includes: 1) Comparison of quantitative indicators: By comparing the evaluation metrics MAE, MSE, and R² of the five models, it was found that the GP-ConvFormer model achieves the best performance in MAE, reducing the error by 35.2% compared to the Transformer; its MSE is only 10.5% of that of SARIMA, and its R² = 0.98 is close to the theoretical limit; moreover, the average prediction time per step of the GP-ConvFormer model is 480ms, which meets the real-time requirements of industrial scenarios. 2) Qualitative visual comparison: By comparing the load prediction performance of five types of models over a six-hour period: SARIMA's prediction curve showed obvious lag and smoothing effects, especially during the start and stop of welding processes; LSTM improved the lag problem, but gradually accumulated biases during continuous operation; CNN could respond quickly to instantaneous changes, but fluctuated too much in long-term predictions; Transformer captured periodic features well, but overshoot occurred at abrupt change points; while GP-ConvFormer performed best overall, accurately following the peak changes of welding torch start and stop, maintaining stable tracking during continuous gas supply, and showing no lag or distortion at any operating condition transition points. Its curve almost coincided with the true value, verifying the advantages of the hybrid network architecture in multi-scale dynamic modeling of shipyard gas load.

8. The method according to claim 7, characterized in that, The prediction results based on the GP-ConvFormer hybrid network model are used for industrial scheduling and energy-saving optimization, specifically including: Based on the multi-regional gas load time-series prediction curves output by the GP-ConvFormer model, and combined with the differences in operating conditions and pressure demand characteristics of each production area, a dynamic scheduling strategy is constructed: By predicting and anticipating peak and off-peak load periods in advance, the number of air compressors started and stopped and their operating power are rationally adjusted to meet the high-pressure requirements of 7.5-8 bar for painting and sandblasting and the process requirements for compensating for instantaneous peak loads on the dock platform, while avoiding redundant operation of multiple devices; In view of the differences in shift collaboration modes and working hours in different regions, the production rhythm patterns captured by the global position bias mechanism are used to achieve coordinated allocation of gas supply resources across regions, reducing pipeline pressure fluctuations and energy waste; In terms of energy-saving optimization, the scheduling of gas storage tanks and the adjustment logic of diversion valves are optimized based on accurate gas load prediction results to reduce the frequency of equipment start-up and shutdown and the idle running time, thereby reducing energy consumption caused by unstable gas supply pressure; At the same time, combined with the unplanned disturbance characteristics identified by the GP-ConvFormer model, maintenance warnings and emergency scheduling plans are triggered in advance to avoid inefficient energy consumption under abnormal operating conditions.

9. A shipyard gas load prediction system based on a hybrid network model, the system comprising a memory for storing computer program instructions and a processor for executing the program instructions, characterized in that, When the computer program instructions are executed by the processor, the system is triggered to execute the shipyard gas load prediction method based on the hybrid network model as described in any one of claims 1 to 8.

Images