An air consumption load prediction method, device, equipment and medium of an air compressor

Abstract

The application relates to the technical field of energy-saving of industrial air compressors, and particularly relates to an air compressor gas load prediction method, device, equipment and medium. In the application, multi-dimensional feature data in multiple historical periods and production data of a to-be-predicted period are input into a trained load prediction model to obtain a predicted load value; when the load prediction model is trained, actual production data is subjected to feature conversion processing to obtain continuous production intensity features; the actual production data and planned production data are subjected to deviation processing to obtain production deviation features; the continuous production intensity features, the production deviation features, equipment gas consumption data and time features are input into the load prediction model for double-weighting processing, and load prediction is carried out based on the weights after the double-weighting processing to obtain a training load value; a loss function is determined based on the training load value and the continuous production intensity features; parameters of the load prediction model are adjusted based on the loss function until a training condition is reached, and the trained load prediction model is obtained.

Description

Technical Field

[0001] This application relates to the field of energy-saving technology for industrial air compressors, and in particular to a method, device, equipment and medium for predicting the air load of an air compressor. Background Technology

[0002] In industrial production, compressed air systems are an important power source, and accurate prediction of their air load is of great significance for achieving energy conservation, consumption reduction, and optimized operation.

[0003] Traditional air compressor load prediction models mainly rely on historical air consumption data for training and prediction. However, the air compressor system's air load is not only affected by historical air consumption data, but also by various factors such as production plans and process changes. Therefore, the prediction results of air compressor load prediction models that rely on historical air consumption data cannot accurately reflect the direct impact of various factors such as production plans and process changes on air demand, thus reducing the accuracy of prediction.

[0004] Therefore, there is an urgent need for a method to predict the air load of air compressors, which is currently not very accurate, in order to improve the accuracy of prediction. Summary of the Invention

[0005] This application provides a method, apparatus, equipment, and medium for predicting the air load of an air compressor, in order to solve the problem of low accuracy in predicting the air load of existing air compressors.

[0006] In a first aspect, this application provides a method for predicting the air load of an air compressor, the method comprising: Acquire multidimensional feature data of the air compressor from multiple historical periods and production data for the period to be predicted; The multidimensional feature data of the multiple historical periods and the production data of the period to be predicted are input into the trained load prediction model to obtain the predicted load value of the period to be predicted. The load forecasting model was trained in the following manner: The process involves acquiring multidimensional feature data from multiple training cycles; this multidimensional feature data includes production data, equipment gas consumption data, and time features; the production data includes actual production data and planned production data; performing feature transformation processing on the actual production data to obtain continuous production intensity features; performing deviation processing on the actual production data and the planned production data to obtain production deviation features; inputting the continuous production intensity features, the production deviation features, the equipment gas consumption data, and the time features into the load prediction model for double weighting processing, and performing load prediction based on the weights after double weighting processing to obtain training load values; determining a loss function based on the training load values ​​and the continuous production intensity features; and adjusting the parameters of the load prediction model based on the loss function until the training conditions are met to obtain a trained load prediction model.

[0007] In one possible implementation, determining the loss function based on the training load value and the continuous production intensity characteristics includes: The baseline prediction error is determined based on the difference between the training load value and the actual load value in the current training cycle. The trend of predicted load change is determined based on the difference between training load values ​​in adjacent training periods; and the trend of continuous production intensity characteristics change is determined based on the difference between continuous production intensity characteristics in adjacent training periods. Based on the predicted load change trend and the continuous production intensity characteristic change trend, a production trend correlation value is constructed. The loss function is determined by weighted summation of the basic prediction error and the production trend correlation value.

[0008] In one possible implementation, determining the trend of predicted load changes based on the difference between training load values ​​in adjacent training cycles includes: Based on the difference between the training load value of the current training cycle and the training load value of the previous training cycle, a first instantaneous change is determined; the direction information of the first instantaneous change is used as the trend of the predicted load change. The determination of the changing trend of the continuous production intensity characteristics based on the differences between the continuous production intensity characteristics of adjacent training periods includes: Based on the difference between the continuous production intensity characteristics of the current training period and the continuous production intensity characteristics of the previous training period, a second instantaneous change is determined; the direction information of the second instantaneous change is used as the changing trend of the continuous production intensity characteristics.

[0009] In one possible implementation, the feature transformation processing of the actual production data to obtain continuous production intensity features includes: The actual production data within the preset sliding window is processed by moving average to obtain a smoothed reference component, which is used to eliminate the step-like jumps caused by discrete counting. Based on the actual production data of the current training cycle and the actual production data of the previous training cycle, the output change rate is calculated; the output change rate is used to characterize the dynamic changes in the production rhythm. The continuous production intensity characteristic is obtained by weighted summation of the production change rate and the smoothed baseline component; the continuous production intensity characteristic is used to reflect the production intensity associated with the gas load change trend.

[0010] In one possible implementation, the deviation processing based on the actual production data and the planned production data to obtain production deviation characteristics includes: Based on the deviation between the actual production data and the planned production data, a planned deviation component is constructed. The rhythm fluctuation component is determined by the standard deviation of the actual production data within a preset sliding window; The production deviation characteristics are obtained by weighted summation of the planned deviation component and the rhythm fluctuation component.

[0011] In one possible implementation, the step of inputting the continuous production intensity characteristics, the production deviation characteristics, the equipment gas consumption data, and the time characteristics into the load forecasting model for dual weighting processing includes: The continuous production intensity characteristics, the production deviation characteristics, the equipment gas consumption data, and the time characteristics are input into the load prediction model, and the weights of each feature are automatically learned through the attention mechanism to obtain the first weight of each feature. Based on the preset feature weighting method, the first weight of each feature is adjusted to obtain the second weight of each feature.

[0012] In one possible implementation, the method further includes: When the predicted load value of the period to be predicted exceeds the preset load range, the moving average method is used to smooth the predicted load value of multiple consecutive prediction periods. The smoothed predicted load value is used as the predicted load value for the period to be predicted.

[0013] Secondly, this application provides an air compressor air load prediction device, the device comprising: The data acquisition module is used to acquire multi-dimensional feature data of the air compressor from multiple historical periods and production data for the period to be predicted. The load forecasting module is used to input the multidimensional feature data of the multiple historical periods and the production data of the period to be predicted into the trained load forecasting model to obtain the predicted load value of the period to be predicted. The model training module is used to train the load prediction model according to the following methods: Acquire multidimensional feature data from multiple training cycles; the multidimensional feature data includes production data, equipment gas consumption data, and time features; the production data includes actual production data and planned production data; perform feature transformation processing on the actual production data to obtain continuous production intensity features; perform deviation processing on the actual production data and the planned production data to obtain production deviation features; input the continuous production intensity features, the production deviation features, the equipment gas consumption data, and the time features into the load prediction model for double weighting processing, and perform load prediction based on the weights after double weighting processing to obtain training load values; determine a loss function based on the training load values ​​and the continuous production intensity features; and adjust the parameters of the load prediction model based on the loss function until the training conditions are met to obtain a trained load prediction model.

[0014] Thirdly, this application also provides an electronic device including a processor, which executes a computer program stored in a memory to implement the steps of a method for predicting the air load of an air compressor as described in any of the first aspects.

[0015] Fourthly, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a method for predicting the air load of an air compressor as described in any of the first aspects.

[0016] Fifthly, this application provides a computer program product, including a computer program: when the computer program is executed by a processor, it implements the steps of a method for predicting the air load of an air compressor as described in any of the first aspects.

[0017] The technical solutions provided by the embodiments of this application have at least the following beneficial effects: This application's embodiments introduce multi-dimensional production-related data to construct continuous production intensity characteristics and production deviation characteristics. When predicting gas load, it can learn the impact of production conditions, thereby improving the accuracy of gas load prediction and its adaptability to actual production conditions. Simultaneously, it applies double weighting to core features strongly correlated with load, increasing the weight of time features and production data in the load prediction model. This allows for more accurate capture of gas consumption patterns, further improving the accuracy of gas load prediction. When calculating the loss function, it focuses on load changes caused by production variations. By penalizing the inconsistency between the predicted load change trend and the production intensity change trend, it forces the model to learn the dynamic coupling relationship between the two, pursuing consistency between the predicted load change trend and the production intensity characteristic change trend, thereby improving the accuracy of gas load prediction. Attached Figure Description

[0018] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application.

[0019] Figure 1 A flowchart illustrating a method for predicting the air load of an air compressor, provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for determining continuous production intensity characteristics provided in an embodiment of this application; Figure 3 A flowchart illustrating a method for determining production deviation characteristics provided in an embodiment of this application; Figure 4 A flowchart illustrating a dual-weighting processing method provided in an embodiment of this application; Figure 5 A flowchart illustrating a loss function determination method provided in an embodiment of this application; Figure 6 A flowchart illustrating the training process of a load forecasting model using production data, as provided in an embodiment of this application. Figure 7 A flowchart illustrating a hybrid prediction model training method provided in an embodiment of this application; Figure 8 A flowchart illustrating a load prediction model training method provided in an embodiment of this application; Figure 9 This is a schematic diagram illustrating a process for actual gas load forecasting, provided as an embodiment of this application. Figure 10 A schematic diagram of the air load prediction device for an air compressor provided in this application embodiment; Figure 11This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The terms "first" and "second" used in this document are for descriptive purposes only and should not be construed as indicating relative importance or implying the number of technical features indicated. In the description of the embodiments of this application, unless otherwise stated, "a plurality of" means two or more.

[0022] In industrial production, compressed air systems are an important power source, and accurate prediction of their air load is of great significance for achieving energy conservation, consumption reduction, and optimized operation.

[0023] Traditional air compressor air load forecasting models mainly rely on historical air consumption data, ignoring production data that is closely related to air load. As a result, the forecast results cannot accurately reflect the direct impact of production conditions such as production plan adjustments and process changes on air demand, thus reducing the accuracy of forecasts.

[0024] The air load of the air compressor system is affected by various factors such as production plans, process changes, and shift changes, exhibiting complex characteristics of nonlinearity, periodic fluctuations, and sudden changes. Traditional air compressor air load prediction models are difficult to effectively capture the complex patterns of data and cannot meet the needs of refined management.

[0025] Traditional air compressor load prediction models perform reasonably well under specific operating conditions, but they are out of sync with production data. This leads to insufficient generalization ability of the models when there are significant changes in the production environment or process, requiring frequent retraining or adjustment of the model parameters.

[0026] Because traditional air compressor load prediction models lack sufficient prediction accuracy and real-time performance, the control logic of air compressor systems often remains at the post-event response stage of "adjusting after pressure fluctuations," making it difficult to effectively predict and optimize future air demand. This restricts the energy-saving potential of air compressor systems and lacks forward-looking optimization capabilities.

[0027] Therefore, there is an urgent need for a method to predict the air load of air compressors, which is currently not very accurate, in order to improve the accuracy of prediction.

[0028] In view of this, this application provides a method, apparatus, equipment and medium for predicting the air load of an air compressor, in order to solve the problem of low accuracy in the prediction of the air load of existing air compressors.

[0029] The inventive concept of this application can be summarized as follows: This application introduces multi-dimensional production-related data to construct continuous production intensity features and production deviation features. When predicting gas load, it can learn the influence of production conditions, thereby improving the accuracy of gas load prediction and its adaptability to actual production conditions. At the same time, it adopts a hybrid prediction model (Attention-LSTM) that embeds a long short-term memory network with an attention mechanism to doubly weight and enhance the core features that are strongly correlated with the load, thereby increasing the weight of time features and production data in the load prediction model. This enables it to capture gas usage patterns more accurately, thereby improving the accuracy of gas load prediction and achieving high-precision prediction.

[0030] This application provides a method for predicting the air load of an air compressor, which can be applied to electronic devices such as PCs, mobile terminals, terminal devices, and servers. Furthermore, this method can be applied to distributed software platforms, such as blockchain.

[0031] See Figure 1 This is a flowchart illustrating a method for predicting the air load of an air compressor according to an embodiment of this application. The method can be specifically executed as follows: Figure 1 The steps shown are as follows: In step S101, multidimensional feature data of multiple historical cycles of the air compressor and production data of the cycle to be predicted are obtained.

[0032] The multidimensional feature data includes production data, equipment gas consumption data, and time characteristics; the production data includes actual production data and planned production data.

[0033] In one possible implementation, to reduce the impact of environmental factors on load forecasting, multidimensional feature data can include environmental features in addition to production data, equipment gas consumption data, and time characteristics. When training the load forecasting model, inputting environmental features allows the model to learn the impact of these features on the load forecasting results. Therefore, during actual forecasting, inputting environmental features into the load forecasting model can more accurately predict load values ​​and reduce the influence of environmental factors on the predicted load values.

[0034] In practice, in order to achieve high-precision prediction, the core of the air load time series data flow of the air compressor system constructed in this application lies in the data fusion and preprocessing of multi-dimensional feature data.

[0035] In order to accurately reflect the direct impact of production conditions such as production plan adjustments and process changes on gas demand, this application has added actual production data and planned production data to its multidimensional feature data, as shown in Table 1.

[0036] Table 1

[0037] In this application, the multidimensional feature data can be data collected in real time or data collected at intervals. For example, all feature data in the multidimensional feature data are time series data with a time granularity of 1 minute.

[0038] To eliminate data noise, dimensional differences, and enhance time period information, the raw data collected in this application embodiment needs to be preprocessed when training the load prediction model and when actually performing load prediction.

[0039] In one possible implementation, the method further includes: normalizing the multidimensional feature data from multiple historical periods and the production data for the period to be predicted.

[0040] Specifically, during the normalization process, min-max normalization is performed on production data, equipment gas consumption data, and environmental characteristics; and sine and cosine periodic transformation is performed on time characteristics.

[0041] In this embodiment, production data (production line output), equipment gas consumption data (gas load, gas supply pressure) and environmental characteristics are processed by min-max normalization. Specifically, the feature values ​​are mapped to the [0, 1] interval to eliminate the difference in dimensions and ensure that the contribution of different features to model training is balanced.

[0042] The continuous data is normalized using the min-max method according to formula (1): (1) in, These are the normalized eigenvalues. These are the original feature values ​​(after preprocessing, with no anomalies or missing values). The minimum value of this feature across all samples. This is the maximum value of the feature across all samples.

[0043] In this embodiment of the application, to address the issue of excessive differences between time features, such as minute-level time data (0-1439 minutes) and equipment feature values, a sine-cosine periodic transformation technique is used to normalize the time features. Specifically, the t-th minute of a day is converted into a two-dimensional periodic feature, effectively guiding the model to capture the daily cycle of gas consumption patterns, while avoiding gradient imbalance caused by differences in dimensions.

[0044] In practice, the time characteristics can be processed by sine and cosine periodic transformation according to formula (2): s (2) Here, sin(t) and cos(t) are two independent input features, which do not require additional normalization processing. It is a time-related feature.

[0045] In step S102, the multidimensional feature data of multiple historical periods and the production data of the period to be predicted are input into the trained load prediction model to obtain the predicted load value of the period to be predicted.

[0046] In this embodiment, the load prediction model employs a hybrid prediction model, Attention-LSTM, which is an Attention mechanism embedded within a Long Short-Term Memory (LSTM) network.

[0047] like Figure 1 As shown in the embodiment of this application, the load prediction model trained in step S102 is obtained by training in the manner described in steps S1031-S1034.

[0048] In step S1031, multidimensional feature data from multiple training cycles are acquired.

[0049] After obtaining the multidimensional feature data, it is necessary to normalize the multidimensional feature data first. The specific normalization method is as described in formulas (1) and (2) above, and will not be elaborated here.

[0050] In step S1032, feature transformation processing is performed on the actual production data to obtain continuous production intensity characteristics; deviation processing is performed based on the actual production data and planned production data to obtain production deviation characteristics.

[0051] Since the original actual production data is usually discrete count values, exhibiting step-like jumps, it does not match the time-series characteristics of continuously fluctuating gas load. Therefore, this application uses mathematical transformation to convert the original discrete actual production data into continuous production intensity characteristics that can reflect continuous load trends.

[0052] In one possible implementation, feature transformation processing is performed on actual production data to obtain continuous production intensity characteristics, which can be executed as follows: Figure 2 The steps shown are as follows: In step S201, the actual production data within the preset sliding window is processed by moving average to obtain a smoothed reference component. The smoothed reference component is used to eliminate the step-like jumps caused by discrete counting.

[0053] In step S202, the output change rate is calculated based on the actual production data of the current training cycle and the actual production data of the previous training cycle; the output change rate is used to characterize the dynamic changes in the production rhythm.

[0054] The production change rate is obtained by dividing the difference between the actual production data of the current training cycle and the actual production data of the previous training cycle by a preset sampling time interval.

[0055] In step S203, the production change rate and the smoothed baseline component are weighted and summed to obtain the continuous production intensity characteristic; the continuous production intensity characteristic is used to reflect the production intensity associated with the gas load change trend.

[0056] During implementation, the original discrete actual production data is first obtained, which consists of discrete count values ​​collected within a preset sampling time interval. The original discrete actual production data is then subjected to continuous feature transformation processing to generate continuous production intensity features.

[0057] In this embodiment of the application, the production change rate and the smoothing benchmark component are weighted and summed according to a first preset weighting coefficient to obtain the continuous production intensity characteristics.

[0058] The first preset weighting coefficient includes a first weighting coefficient and a second weighting coefficient. The first weighting coefficient corresponds to the smoothed baseline component, and the second weighting coefficient corresponds to the rate of change in output. The sum of the first weighting coefficient and the second weighting coefficient is 1, and the first weighting coefficient is greater than the second weighting coefficient.

[0059] The first preset weighting coefficient is used to balance the production baseline level and changes in production rhythm. Therefore, the value of the first weighting coefficient is related to the smoothness of the production line operation. When the production line is running smoothly, the first weighting coefficient is increased to obtain a smoother trend and enhance the contribution of the smoothing baseline component; when the production line starts and stops frequently, the first weighting coefficient is decreased to enhance the capture of dynamic changes and enhance the contribution of the output change rate.

[0060] By using continuous production intensity features as input features to train the Attention-LSTM hybrid prediction model, the model learns the mapping relationship between production intensity and gas load, thereby further improving the accuracy of the model's predictions.

[0061] For example, the continuous production intensity characteristics can be obtained by performing feature transformation processing on the actual production data using formula (3): (3) in, for The continuous production intensity characteristics at any given moment; To smooth the baseline component, i.e., the actual production data within a preset sliding window The moving average within the range is used to eliminate the stair-step effect of discrete counts; for Raw, discrete, actual production data collected in real time; The preset sampling time interval; for The raw, discrete, actual production data collected in the previous training cycle at that time.

[0062] This is the rate of change in output, used to capture the dynamic changes in the production line during start-up, shutdown, or speed-up. The first weighting coefficient corresponding to the smoothing reference component is, in this application The value range can be set to [0.6, 0.8], or it can be set according to actual needs. This is the second weighting coefficient corresponding to the rate of change in output.

[0063] Formula (3) maps discrete actual production data into production intensity characteristics with continuous time series features. During the optimization process, if the predicted load is slow to respond to production fluctuations, the load should be reduced. To increase the weight of the rate of change in output.

[0064] In actual operation, the deviation between "actual output" and "planned output" will affect the gas load. Therefore, during training, the model needs to learn the impact of unexpected production changes on the load by analyzing the deviation between actual production data and planned production data.

[0065] To address this, this application constructs a dynamic response feature based on the "planned-actual" production deviation, namely, the production deviation feature. This feature is introduced because the model needs to learn the complex mapping relationship between actual production data and gas load during the training phase, while the prediction phase relies on future planned production data. However, due to unexpected production changes, deviations often occur between actual and planned production data. Therefore, it is necessary to learn the impact of the production deviation feature on gas load.

[0066] In one possible implementation, production deviation features are used to characterize production instability; these features include a plan deviation component and a rhythm fluctuation component. Therefore, by processing deviations based on actual production data and planned production data, production deviation features are obtained, which can be implemented as follows: Figure 3 The steps shown are as follows: In step S301, a planned deviation component is constructed based on the deviation between actual production data and planned production data.

[0067] In this application, the planned deviation component is constructed based on the absolute value of the difference between actual production data and planned production data. The planned deviation component is used to reflect the degree of deviation of actual production data from planned production data.

[0068] In step S302, the rhythm fluctuation component is determined by the standard deviation of the actual production data within a preset sliding window.

[0069] In this application, the rhythm fluctuation component is based on actual production data within a preset sliding window. The standard deviation within the range is constructed. The rhythm fluctuation component is used to reflect the dispersion of production rhythm.

[0070] In step S303, the planned deviation component and the rhythm fluctuation component are weighted and summed to obtain the production deviation characteristics.

[0071] In this embodiment of the application, the planned deviation component and the rhythm fluctuation component are weighted and summed according to a second preset weighting coefficient to obtain the production deviation characteristics.

[0072] The second preset weighting coefficient includes a third weighting coefficient and a fourth weighting coefficient. The third weighting coefficient corresponds to the plan deviation component, and the fourth weighting coefficient corresponds to the rhythm fluctuation component. The sum of the third and fourth weighting coefficients is 1. The second preset weighting coefficient is used to balance the contribution of the plan deviation component and the rhythm fluctuation component to the production deviation characteristics.

[0073] Among them, the planned deviation component and the rhythm fluctuation component are normalized by the maximum output reference benchmark to ensure that the production deviation characteristics are consistent across different production line scales.

[0074] By using production deviation features as input features to train the Attention-LSTM hybrid prediction model, the model learns the impact of unexpected production changes on gas load, thereby further improving the accuracy of the model's predictions.

[0075] For example, the deviation between actual production data and planned production data can be processed using formula (4) to obtain the production deviation characteristics: (4) in, for The production deviation characteristic at any given time is a dimensionless quantity used to characterize production instability; This refers to the deviation component from the plan, i.e., actual production data. With planned production data The absolute value of the difference; This refers to the rhythm fluctuation component, i.e., the actual production data. In the preset sliding window The standard deviation within the range is used to quantify the dispersion of production rhythm; It serves as a reference benchmark for maximum output and is used to uniformly normalize the components of plan deviation and rhythm fluctuation, so that the model can be consistent under different production line scales. To find the minimum value, to prevent the denominator from being zero; The third weighting coefficient, It is the fourth weighting coefficient, used to balance the contribution of the plan deviation component and the rhythm fluctuation component to the production deviation characteristics. and The value range can be set to [0.3, 0.7], or it can be set according to actual needs. .

[0076] In step S1033, the continuous production intensity characteristics, production deviation characteristics, equipment gas consumption data and time characteristics are input into the load prediction model for double weighting processing, and load prediction is performed based on the weights after double weighting processing to obtain the training load value.

[0077] In one possible implementation, continuous production intensity characteristics, production deviation characteristics, equipment gas consumption data, and time characteristics are input into the load forecasting model and subjected to double weighting processing, which can be executed as follows: Figure 4 The steps shown are as follows: In step S401, the continuous production intensity characteristics, production deviation characteristics, equipment gas consumption data and time characteristics are input into the load prediction model, and the weights of each feature are automatically learned through the attention mechanism to obtain the first weight of each feature. In step S402, based on the preset feature weighting method, the first weight of each feature is adjusted to obtain the second weight of each feature.

[0078] In practice, this application inputs continuous production intensity characteristics, production deviation characteristics, equipment gas consumption data and time characteristics into the load prediction model, and then performs double weighting on each input characteristic in the load prediction model to guide the model to focus on core characteristics and avoid interference from weak characteristics.

[0079] Dual weighting processing includes automatic weighting via the Attention mechanism and manual weighting.

[0080] The Attention mechanism automatically weights features by learning the weights of each feature and time step through the attention mechanism of the Attention layer, thereby achieving adaptive feature importance allocation and obtaining the first weight of each feature.

[0081] Manual weighting involves pre-configuring a static feature weighting method based on prior knowledge. After obtaining the first weight of each feature, the first weight of each target feature is adjusted again using the pre-configured feature weighting method to obtain the second weight of each target feature, thereby further strengthening the dominant role of each target feature in the prediction result.

[0082] The target feature is the feature whose correlation with gas load is greater than a preset correlation threshold, such as features with high correlation to load, historical gas load, and production line output parameters. The target feature can be all or some of the features.

[0083] In step S1034, a loss function is determined based on the training load value and continuous production intensity characteristics; and the parameters of the load prediction model are adjusted based on the loss function until the training conditions are met, thus obtaining a completed load prediction model.

[0084] In one possible implementation, the loss function is determined based on the training load value and continuous production intensity characteristics, which can be performed as follows: Figure 5 The steps shown are as follows: In step S501, the basic prediction error is determined based on the difference between the training load value and the actual load value in the current training cycle.

[0085] Among them, the basic prediction error is used to constrain the model's performance in numerical prediction accuracy.

[0086] In step S502, the trend of predicted load change is determined based on the difference between training load values ​​in adjacent training periods; and the trend of continuous production intensity characteristics is determined based on the difference between continuous production intensity characteristics in adjacent training periods.

[0087] In one possible implementation, determining the trend of predicted load change based on the difference between training load values ​​in adjacent training periods includes: Based on the difference between the training load value of the current training cycle and the training load value of the previous training cycle, the first instantaneous change is determined; the direction information of the first instantaneous change is used as the predicted trend of load change.

[0088] In one possible implementation, determining the changing trend of continuous production intensity characteristics based on the differences between continuous production intensity characteristics in adjacent training periods includes: Based on the difference between the continuous production intensity characteristics of the current training period and the continuous production intensity characteristics of the previous training period, the second instantaneous change is determined; the directional information of the second instantaneous change is used as the trend of the continuous production intensity characteristics.

[0089] The directional information of the first instantaneous change and the directional information of the second instantaneous change are determined by a sign function, which is used to characterize the rising, falling or unchanged state of the instantaneous change.

[0090] In step S503, a production trend correlation value is constructed based on the predicted load change trend and the continuous production intensity characteristic change trend.

[0091] Among them, the production trend correlation value is constructed based on the consistency between the trend of predicted load change and the trend of continuous production intensity characteristics, and is used to penalize the inconsistency between the direction of predicted load change and the direction of continuous production intensity characteristics change.

[0092] In one possible implementation, a production trend correlation value is constructed based on the predicted load change trend and the continuous production intensity characteristic change trend, and the following steps are performed: Calculate the difference between the directional information of the first instantaneous change and the directional information of the second instantaneous change; sum and average the difference over all training periods to obtain the production trend correlation value.

[0093] In step S504, the basic prediction error and the production trend correlation value are weighted and summed to determine the loss function.

[0094] This loss function is used to guide the model to learn the dynamic coupling relationship between production change trends and gas load change trends during the model training process.

[0095] In this embodiment, the production trend correlation value is multiplied by a preset production correlation weight coefficient and then added to the basic prediction error to obtain the loss function.

[0096] The preset production correlation weighting coefficient is used to balance the contribution of the basic forecast error and the production trend correlation value to the total loss. Its value range can be set to [0.1, 0.5], or it can be set according to actual needs.

[0097] During the training of the load forecasting model, the goal is to minimize the loss function, so that while maintaining the accuracy of numerical forecasting, the load forecasting is forced to learn the dynamic consistency between the trend of load change and the trend of continuous production intensity characteristics.

[0098] For example, the loss function can be determined using formula (5): (5) in, The loss function is used to train the model, and the training objective is to minimize the value of the loss function. Based on the prediction error, used to ensure the training load value. The numerical value should be as close as possible to the actual load value. ; To predict the first instantaneous change in load; This represents the second instantaneous change in the intensity characteristics of continuous production; For a sign function, when hour When it is a positive number, hour When it is negative, hour The sign value is 0. This sign function is used to extract the direction information of instantaneous changes; The total number of training cycles; These are preset production-related weighting coefficients.

[0099] By incorporating production trend correlation values ​​into the loss function, the load forecasting model is guided to focus on load changes caused by production fluctuations during training. Through this loss function, the load forecasting model not only strives for numerical approximation during learning but also seeks consistency between the predicted load change trend and the continuous production intensity characteristic change trend. This application forces the load forecasting model to learn the dynamic coupling relationship between the two by penalizing the inconsistency between the predicted load change trend and the continuous production intensity characteristic change trend.

[0100] A flowchart illustrating the application of production data in the training process of a load forecasting model is shown below. Figure 6 As shown. In the model training phase, this application uses actual production data as the core dimension of model learning. Through multi-step feature mapping and correlation modeling, the load forecasting model deeply understands the coupling relationship between production behavior and gas load.

[0101] like Figure 6 As shown, the actual production data is first processed through feature transformation to obtain continuous production intensity features. Then, deviation processing is performed on the actual and planned production data to obtain production deviation features. These features, along with continuous production intensity features, production deviation features, equipment gas consumption data, time features, and environmental features, are then used to construct a multi-source production data feature vector. This multi-source production data feature vector is input into the load forecasting model for training. A loss function is determined based on the training load values ​​and the introduced production trend correlation values. If the accuracy of the load forecasting model is insufficient (i.e., the training conditions are not met), the parameters and weights of the load forecasting model are adjusted based on the loss function, and the model is trained again until the accuracy meets the requirements. The trained load forecasting model is then obtained and saved.

[0102] In the embodiments of this application, the training condition can be either minimizing the loss function of the load prediction model or reaching a threshold number of training iterations.

[0103] In one possible implementation, the Attention-LSTM hybrid prediction model used in this application introduces an attention mechanism at the model's encoding layer, allowing the model to automatically learn the importance of different time steps in the historical time series, assigning high weights to information from key time steps and low weights to irrelevant information. Its core principle is reflected in the feature weighting mechanism.

[0104] like Figure 7 The diagram illustrates the training process of the Attention-LSTM hybrid prediction model. This model comprises an input layer, an LSTM layer, an Attention layer, and an output layer. The input layer standardizes the multidimensional feature data. The LSTM layer processes temporal features and time dependencies. The Attention layer automatically learns the weights of each feature and time step, achieving adaptive feature importance allocation. Then, for parameters highly correlated with load, such as time-related parameters, historical gas load, and production line output, a manual weighting method is used to further strengthen their dominant role in the prediction results. This increases the weight of time features and production data in the load prediction model, enabling more accurate capture of gas usage patterns and thus improving the accuracy of gas load prediction, achieving high-precision prediction. The output layer outputs the predicted load value.

[0105] This application's embodiments implement multiple training optimization strategies in the hybrid prediction model Attention-LSTM to achieve high-precision prediction of air compressor air load. For example... Figure 7 As shown, an early stopping strategy and the ReduceLROnPlateau learning rate scheduler are introduced to dynamically adjust the learning rate, thereby solving the problem of model overfitting.

[0106] like Figure 7 As shown in the embodiment of this application, after stopping training and saving the model parameters, the load prediction model is evaluated using core indicators such as MAPE (Mean Absolute Percentage Error) and auxiliary indicators such as R² (Coefficient of Determination). If the indicators meet the standards, the trained load prediction model is obtained. If the indicators do not meet the standards, the weights of each feature and the hyperparameters of the model are adjusted, and the load prediction model is retrained based on the adjusted weights of each feature and the hyperparameters of the model.

[0107] MAPE is calculated as the average of "absolute error / true value" and presented as a percentage. It is a core indicator for evaluating forecast accuracy and one of the most practical evaluation indicators in time series forecasting.

[0108] The auxiliary indicator R² (coefficient of determination) is used to measure the model's ability to capture changes in the target variable (instantaneous flow rate), reflecting the proportion of variance (fluctuation) of the target variable explained by the model, i.e., the model's goodness of fit.

[0109] Combination Figure 6 and Figure 7 The training process, obtained Figure 8 The training process for the load forecasting model is shown below. First, feature transformation is performed on the actual production data to obtain continuous production intensity features. Then, deviation processing is performed on the actual production data and planned production data to obtain production deviation features. Next, continuous production intensity features, production deviation features, equipment gas consumption data, time features, and environmental features are used to construct a multi-source production data feature vector. This multi-source production data feature vector is then input into the load forecasting model.

[0110] The input layer standardizes the multidimensional feature data; the LSTM layer processes temporal features and time dependencies; the Attention layer performs double weighting to increase the weight of temporal features and production data in the load prediction model, and load prediction is performed based on the double-weighted weights to obtain the training load value. The predicted training load value is output at the output layer.

[0111] Then, the loss function is determined based on the training load value output by the output layer and the introduced production trend correlation value. If the accuracy of the load prediction model is insufficient, the learning rate is dynamically adjusted and the model is retrained; if the accuracy of the load prediction model is sufficient, training is stopped and the model parameters are saved; then, indicators such as MAPE are calculated; if the indicators are satisfactory, the model parameters are saved, and the completed load prediction model is trained. If the indicators are insufficient, the weights of each feature and the model's hyperparameters are adjusted, and the load prediction model is retrained based on the adjusted weights of each feature and the model's hyperparameters.

[0112] Therefore, by making full use of production data during the model training process, the dynamics and accuracy of the model's prediction of gas load have been improved.

[0113] In actual gas load forecasting, a linkage mechanism of "real-time data-driven + dynamic closed-loop optimization" is constructed to ensure that the forecast output can keep up with changes in production site data. The model's gas load forecasting process is as follows: Figure 9 As shown.

[0114] First, actual production data for the period to be predicted is obtained from sensors on the production site and the production management system. Then, planned production data for the period to be predicted is obtained from the production planning system.

[0115] All actual and planned production data are stored in the production database. When forecasting gas load, multi-dimensional feature data from multiple historical periods and actual and planned production data for the period to be predicted are retrieved from the production database in real time and sent to the hybrid forecasting model Attention-LSTM for load forecasting, and the load forecasting results are output.

[0116] Multidimensional feature data from multiple historical periods can be accessed as needed. For example, if a hybrid prediction model is automatically trained once a week, then a 7-day update cycle is used, and multiple prediction cycles preceding the period to be predicted within that update cycle are considered as multiple historical periods.

[0117] For example, if the period to be predicted is Thursday, then the multidimensional feature data for multiple historical periods includes the multidimensional feature data for Monday, Tuesday, and Wednesday; if the period to be predicted is Saturday, then the multidimensional feature data for multiple historical periods includes the multidimensional feature data for Monday, Tuesday, Wednesday, Thursday, and Friday.

[0118] In this embodiment, actual production data from the production line is synchronously updated and stored in the production database via platform integration or sensor acquisition. When the prediction engine initiates a prediction request, it directly and instantly retrieves the latest actual production data and planned production data for the period to be predicted from the production database.

[0119] In the embodiments of this application, all input features of the load forecasting model, including load features and production data, are real-time online data to ensure that the forecasting benchmark is fully aligned with the current production status.

[0120] In one possible implementation, the method further includes: The system automatically trains the completed load forecasting model according to the update cycle, fine-tunes the parameters of the completed load forecasting model, and uses the fine-tuned load forecasting model as the new completed load forecasting model.

[0121] like Figure 9 As shown, with the update cycle set to once a week, the latest real-time production data continuously collected from the production database is retrieved weekly. Based on this data, the load forecasting model is automatically trained, dynamically modifying model parameters and feature weights. Fine-tuning of parameters such as model parameters and weight allocation is performed, and the adjusted load forecasting model replaces the existing model in the production database. This avoids model oscillations caused by large-scale parameter adjustments while ensuring accurate matching of parameter adjustments to changes in production rhythm. Through iterative automatic training and parameter fine-tuning, the load forecasting model continuously learns the dynamic changes in production rhythm, improving its adaptability to complex production environments and operating conditions.

[0122] In one possible implementation, the method further includes: When the predicted load value of the forecast period exceeds the preset load range, the moving average method is used to smooth the predicted load values ​​of multiple consecutive forecast periods. The smoothed forecast load value is used as the forecast load value for the period to be forecasted.

[0123] Among them, multiple consecutive prediction periods include the period to be predicted and the n prediction periods preceding the period to be predicted.

[0124] Specifically, when the predicted value exceeds the reasonable range, the predicted value output by the model is corrected to ensure that the predicted value is within the actual load range of the air compressor system. As shown in formula (6), assuming a prediction cycle is 1 minute and the prediction cycle is 2, the moving average method is used to smooth the predicted values ​​of three consecutive prediction cycles to reduce prediction fluctuations and improve the stability of the prediction results, while the model optimization process is started.

[0125] (6)

[0126] in, This represents the predicted load value after smoothing at minute t. , , These are the predicted load values ​​for minutes t-2, t-1, and t, respectively.

[0127] Therefore, this application introduces multi-dimensional production-related data to construct continuous production intensity characteristics and production deviation characteristics. When predicting gas load, it can learn the impact of production conditions, thereby improving the accuracy of gas load prediction and its adaptability to actual production conditions. Simultaneously, it applies double weighting to core features strongly correlated with load, increasing the weight of time features and production data in the load prediction model. This allows for more accurate capture of gas consumption patterns, further improving the accuracy of gas load prediction. When calculating the loss function, it focuses on load changes caused by production variations. By penalizing the inconsistency between the predicted load change trend and the production intensity change trend, it forces the model to learn the dynamic coupling relationship between the two, pursuing consistency between the predicted load change trend and the production intensity characteristic change trend, thereby improving the accuracy of gas load prediction.

[0128] Based on the same inventive concept, this application also provides an air compressor air load prediction device. Figure 10 A schematic diagram of an air compressor air load prediction device provided in this application embodiment is shown. The device includes: The data acquisition module 1001 is used to acquire multi-dimensional feature data of the air compressor for multiple historical periods and production data for the period to be predicted; The load forecasting module 1002 is used to input the multidimensional feature data of the multiple historical periods and the production data of the period to be predicted into the trained load forecasting model to obtain the predicted load value of the period to be predicted. Model training module 1003 is used to train the load prediction model according to the following method: Acquire multidimensional feature data from multiple training cycles; the multidimensional feature data includes production data, equipment gas consumption data, and time features; the production data includes actual production data and planned production data; perform feature transformation processing on the actual production data to obtain continuous production intensity features; perform deviation processing on the actual production data and the planned production data to obtain production deviation features; input the continuous production intensity features, the production deviation features, the equipment gas consumption data, and the time features into the load prediction model for double weighting processing, and perform load prediction based on the weights after double weighting processing to obtain training load values; determine a loss function based on the training load values ​​and the continuous production intensity features; and adjust the parameters of the load prediction model based on the loss function until the training conditions are met to obtain a trained load prediction model.

[0129] In one possible implementation, the model training module 1003 is specifically used for: The baseline prediction error is determined based on the difference between the training load value and the actual load value in the current training cycle. The trend of predicted load change is determined based on the difference between training load values ​​in adjacent training periods; and the trend of continuous production intensity characteristics change is determined based on the difference between continuous production intensity characteristics in adjacent training periods. Based on the predicted load change trend and the continuous production intensity characteristic change trend, a production trend correlation value is constructed. The loss function is determined by weighted summation of the basic prediction error and the production trend correlation value.

[0130] In one possible implementation, the model training module 1003 is specifically used for: Based on the difference between the training load value of the current training cycle and the training load value of the previous training cycle, a first instantaneous change is determined; the direction information of the first instantaneous change is used as the trend of the predicted load change. The determination of the changing trend of the continuous production intensity characteristics based on the differences between the continuous production intensity characteristics of adjacent training periods includes: Based on the difference between the continuous production intensity characteristics of the current training period and the continuous production intensity characteristics of the previous training period, a second instantaneous change is determined; the direction information of the second instantaneous change is used as the changing trend of the continuous production intensity characteristics.

[0131] In one possible implementation, the model training module 1003 is specifically used for: The actual production data within the preset sliding window is processed by moving average to obtain a smoothed reference component, which is used to eliminate the step-like jumps caused by discrete counting. Based on the actual production data of the current training cycle and the actual production data of the previous training cycle, the output change rate is calculated; the output change rate is used to characterize the dynamic changes in the production rhythm. The continuous production intensity characteristic is obtained by weighted summation of the production change rate and the smoothed baseline component; the continuous production intensity characteristic is used to reflect the production intensity associated with the gas load change trend.

[0132] In one possible implementation, the model training module 1003 is specifically used for: Based on the deviation between the actual production data and the planned production data, a planned deviation component is constructed. The rhythm fluctuation component is determined by the standard deviation of the actual production data within a preset sliding window; The production deviation characteristics are obtained by weighted summation of the planned deviation component and the rhythm fluctuation component.

[0133] In one possible implementation, the model training module 1003 is specifically used for: The continuous production intensity characteristics, the production deviation characteristics, the equipment gas consumption data, and the time characteristics are input into the load prediction model, and the weights of each feature are automatically learned through the attention mechanism to obtain the first weight of each feature. Based on the preset feature weighting method, the first weight of each feature is adjusted to obtain the second weight of each feature.

[0134] In one possible implementation, the apparatus further includes a result optimization module 1004; the result optimization module 1004 is specifically used for: When the predicted load value of the period to be predicted exceeds the preset load range, the moving average method is used to smooth the predicted load value of multiple consecutive prediction periods. The smoothed predicted load value is used as the predicted load value for the period to be predicted.

[0135] Based on the same inventive concept, this application also provides an electronic device. Figure 11 This application provides a schematic diagram of an electronic device structure, such as... Figure 11 As shown, it includes: processor 1101, communication interface 1102, memory 1103 and communication bus 1104, wherein processor 1101, communication interface 1102 and memory 1103 communicate with each other through communication bus 1104. The memory 1103 stores a computer program. When the program is executed by the processor 1101, the processor 1101 performs the steps of any of the air compressor air load prediction methods provided in the embodiments of this application.

[0136] Since the technical problem solved by the above-mentioned electronic device is similar to that of a method for predicting the air load of an air compressor, the implementation of the above-mentioned electronic device can be found in the embodiments of the method, and repeated details will not be described again.

[0137] The communication bus mentioned in the aforementioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used in the figure, but this does not indicate that there is only one bus or one type of bus. Communication interface 1102 is used for communication between the aforementioned electronic device and other devices. The memory can include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0138] The processors mentioned above can be general-purpose processors, including central processing units, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits, field-programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0139] Based on the above embodiments, this invention also provides a computer-readable storage medium storing a computer program executable by a processor. When the program runs on the processor, it causes the processor to perform the steps of any of the air compressor air load prediction methods provided in this application.

[0140] Based on the same inventive concept, embodiments of this application provide a computer program product, including a computer program, which, when executed by a processor, implements the steps of any of the air compressor air load prediction methods provided in embodiments of this application.

[0141] Since the principle of the computer-readable storage medium in solving the problem is similar to that of a method for predicting the air load of an air compressor, the implementation of the computer-readable storage medium can be found in the embodiments of the method, and repeated details will not be described again.

[0142] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0143] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0144] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0145] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0146] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for predicting the air load of an air compressor, characterized in that, The method includes: Acquire multidimensional feature data of the air compressor from multiple historical periods and production data for the period to be predicted; The multidimensional feature data of the multiple historical periods and the production data of the period to be predicted are input into the trained load prediction model to obtain the predicted load value of the period to be predicted. The load forecasting model was trained in the following manner: Acquire multidimensional feature data from multiple training cycles; the multidimensional feature data includes production data, equipment gas consumption data, and time features; the production data includes actual production data and planned production data; perform feature transformation processing on the actual production data to obtain continuous production intensity features; perform deviation processing on the actual production data and the planned production data to obtain production deviation features; input the continuous production intensity features, the production deviation features, the equipment gas consumption data, and the time features into the load prediction model for double weighting processing, and perform load prediction based on the weights after double weighting processing to obtain training load values; determine the training load value based on the training load value and the continuous production intensity features. A loss function is defined; and the parameters of the load forecasting model are adjusted based on the loss function until the training conditions are met, resulting in a trained load forecasting model. The step of inputting the continuous production intensity features, production deviation features, equipment gas consumption data, and time features into the load forecasting model for dual weighting includes: inputting the continuous production intensity features, production deviation features, equipment gas consumption data, and time features into the load forecasting model; automatically learning the weights of each feature through an attention mechanism to obtain the first weight of each feature; and adjusting the first weight of each feature based on a preset feature weighting method to obtain the second weight of each feature.

2. The method according to claim 1, characterized in that, The determination of the loss function based on the training load value and the continuous production intensity characteristics includes: The baseline prediction error is determined based on the difference between the training load value and the actual load value in the current training cycle. The trend of predicted load change is determined based on the difference between training load values ​​in adjacent training periods; and the trend of continuous production intensity characteristics change is determined based on the difference between continuous production intensity characteristics in adjacent training periods. Based on the predicted load change trend and the continuous production intensity characteristic change trend, a production trend correlation value is constructed. The loss function is determined by weighted summation of the basic prediction error and the production trend correlation value.

3. The method according to claim 2, characterized in that, Determining the trend of predicted load changes based on the difference between training load values ​​in adjacent training cycles includes: Based on the difference between the training load value of the current training cycle and the training load value of the previous training cycle, a first instantaneous change is determined; the direction information of the first instantaneous change is used as the trend of the predicted load change. The determination of the changing trend of the continuous production intensity characteristics based on the differences between the continuous production intensity characteristics of adjacent training periods includes: Based on the difference between the continuous production intensity characteristics of the current training period and the continuous production intensity characteristics of the previous training period, a second instantaneous change is determined; the direction information of the second instantaneous change is used as the changing trend of the continuous production intensity characteristics.

4. The method according to claim 1, characterized in that, The step of performing feature transformation processing on the actual production data to obtain continuous production intensity features includes: The actual production data within the preset sliding window is processed by moving average to obtain a smoothed reference component, which is used to eliminate the step-like jumps caused by discrete counting. Based on the actual production data of the current training cycle and the actual production data of the previous training cycle, the output change rate is calculated; the output change rate is used to characterize the dynamic changes in the production rhythm. The continuous production intensity characteristic is obtained by weighted summation of the production change rate and the smoothed baseline component; the continuous production intensity characteristic is used to reflect the production intensity associated with the gas load change trend.

5. The method according to claim 1, characterized in that, The deviation processing based on the actual production data and the planned production data to obtain production deviation characteristics includes: Based on the deviation between the actual production data and the planned production data, a planned deviation component is constructed. The rhythm fluctuation component is determined by the standard deviation of the actual production data within a preset sliding window; The production deviation characteristics are obtained by weighted summation of the planned deviation component and the rhythm fluctuation component.

6. The method according to claim 1, characterized in that, The method further includes: When the predicted load value of the period to be predicted exceeds the preset load range, the moving average method is used to smooth the predicted load value of multiple consecutive prediction periods. The smoothed predicted load value is used as the predicted load value for the period to be predicted.

7. A device for predicting the air load of an air compressor, characterized in that, The device includes: The data acquisition module is used to acquire multi-dimensional feature data of the air compressor from multiple historical periods and production data for the period to be predicted. The load forecasting module is used to input the multidimensional feature data of the multiple historical periods and the production data of the period to be predicted into the trained load forecasting model to obtain the predicted load value of the period to be predicted. The model training module is used to train the load prediction model according to the following methods: The system acquires multidimensional feature data from multiple training cycles; the multidimensional feature data includes production data, equipment gas consumption data, and time features; the production data includes actual production data and planned production data; it performs feature transformation processing on the actual production data to obtain continuous production intensity features; it performs deviation processing on the actual production data and the planned production data to obtain production deviation features; it inputs the continuous production intensity features, the production deviation features, the equipment gas consumption data, and the time features into the load prediction model for double weighting processing, and performs load prediction based on the weights after double weighting processing to obtain training load values; it determines a loss function based on the training load values ​​and the continuous production intensity features; and it adjusts the parameters of the load prediction model based on the loss function until the training conditions are met to obtain a trained load prediction model; the model training module is specifically used for: inputting the continuous production intensity features, the production deviation features, the equipment gas consumption data, and the time features into the load prediction model, automatically learning the weights of each feature through an attention mechanism to obtain the first weight of each feature; and adjusting the first weights of each feature based on a preset feature weighting method to obtain the second weights of each feature.

8. An electronic device, characterized in that, The electronic device includes a processor that executes a computer program stored in a memory to implement the steps of the air compressor air load prediction method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of a method for predicting the air load of an air compressor as described in any one of claims 1-6.

Images