A heat transfer process multi-period energy efficiency prediction method fusing prior knowledge and related device

Abstract

The application provides a heat transfer process multi-period energy efficiency prediction method fusing prior knowledge and a related device, and relates to the technical field of energy efficiency prediction. Based on the prior knowledge of the fouling process, a fouling data set is divided into a cumulative data set and a sliding window data set, a multi-period parallel cascade ConBiLSTM model is constructed, including a CNN module and a BiLSTM module, a preset event trigger condition is acquired, when it is determined that the preset event trigger condition is met, the cumulative data set and the sliding window data set are dynamically updated through an adaptive online learning mechanism; initial prediction is performed through the CNN module; the CNN module and the BiLSTM module are retrained according to the updated data set, and the final fouling resistance prediction value is generated by combining the outputs of the two; and the maintenance scheduling strategy of the heat transfer system is optimized based on the prediction result. The technical effect of accurately predicting the fouling behavior is realized.

Description

Technical Field

[0001] This application relates to the field of energy efficiency prediction technology, and in particular to a multi-cycle energy efficiency prediction method and related apparatus for heat transfer processes that integrates prior knowledge. Background Technology

[0002] Energy efficiency prediction has always been a critical and challenging task in heat transfer processes, and is essential for energy recovery in process industries. Scale buildup leads to reduced heat transfer efficiency, increased energy consumption, operational interruptions, and higher maintenance costs. Accurate and reliable energy efficiency prediction is crucial for optimizing the structural design of heat transfer processes, developing intelligent control systems, and formulating effective energy-saving strategies and maintenance plans. However, existing prediction models often have limitations in handling the multi-cycle characteristics of heat transfer efficiency, especially the impact of cleaning operations, which can cause sudden resets in heat transfer efficiency, resulting in unstable changes in energy efficiency. Therefore, there is an urgent need for advanced models that can integrate spatiotemporal characteristics and domain knowledge to better capture the complex dynamics of heat transfer efficiency across different cycles, particularly during cleaning events. Summary of the Invention

[0003] The core of this invention lies in providing a multi-cycle energy efficiency prediction method and related device for heat transfer processes that integrates prior knowledge, thereby solving the problem of difficulty in predicting the complex dynamics of heat transfer efficiency in different cycles in the prior art.

[0004] Firstly, the multi-cycle energy efficiency prediction method for heat transfer processes that integrates prior knowledge, provided in this application, adopts the following technical solution:

[0005] A multi-cycle energy efficiency prediction method for heat transfer processes that integrates prior knowledge includes:

[0006] Based on the prior knowledge of the influence of scaling on heat transfer efficiency and the formation of scaling, the heat transfer efficiency dataset is divided into a cumulative dataset and a sliding window dataset. The cumulative dataset is used to model spatial nonlinear characteristics, and the sliding window dataset is used to capture short-term time dynamics.

[0007] A multi-cycle parallel cascaded ConBiLSTM model is constructed, including a CNN module and a BiLSTM module, wherein the CNN module is used to extract the spatial dependencies of process variables, and the BiLSTM module is used to model the bidirectional temporal correlation of heat transfer efficiency.

[0008] Obtain preset event trigger conditions, including prediction error trigger conditions and cleaning operation trigger conditions, and start model update when any condition is met;

[0009] When the preset event triggering conditions are met, the accumulated dataset and the sliding window dataset are dynamically updated through an adaptive online learning mechanism;

[0010] After the cleaning event, the prediction of the BiLSTM module is paused, and the initial prediction is made only by the CNN module until the sliding window dataset accumulates at least H new samples;

[0011] The CNN module and BiLSTM module are retrained based on the updated dataset, and the outputs of both are combined to generate the final predicted heat transfer efficiency.

[0012] Optimize the maintenance scheduling strategy of the heat transfer system based on the prediction results.

[0013] Optionally, in the step of dividing the heat transfer efficiency dataset into a cumulative dataset and a sliding window dataset based on prior knowledge of the influence of scaling on heat transfer efficiency and scaling formation, the cumulative dataset X... (s) Through formula

[0014]

[0015] Aggregate historical data, where X (i) Let be the process variable for stage i, and s represent stage s.

[0016] Optionally, the kernel width of the CNN module is 1 to 7, used to simultaneously extract spatial features of adjacent process variables. In the s-th stage, the convolution operation is as follows:

[0017]

[0018] in, It is the activation value of the k-th convolutional kernel at position i. The input vector X (s) In the (i+j)th variable, h represents the width. It is the weight of the j-th element in the k-th convolutional kernel, b (k) σ is the bias term, and σ(·) is the activation function.

[0019] Optionally, the step of obtaining preset event triggering conditions, including prediction error triggering conditions and cleaning operation triggering conditions, and initiating model updates when either condition is met, includes:

[0020] Define a prediction error trigger condition: when the prediction error within the historical window exceeds a threshold δ1, an update is initiated.

[0021]

[0022] in, It is the prediction error, μ j is the weight of each error term, and h is the size of the historical window;

[0023] Define the triggering condition for the cleaning operation. When a cleaning operation (t) is detected... k =t clean The update is triggered when )

[0024]

[0025] The preset event triggering conditions are expressed as follows:

[0026]

[0027] Optionally, the step of dynamically updating the accumulated dataset and the sliding window dataset through an adaptive online learning mechanism includes:

[0028] After determining that the preset event triggering condition is met. k At that time, the sliding window dataset Y is reset through an adaptive online learning mechanism. (s) The sliding window needs to collect H new samples before it can continue making predictions.

[0029]

[0030] Cumulative dataset X (s) Unaffected by cleaning and resetting, the time dataset Y generated during the heat transfer efficiency cycle. (s) The formula for calculating the total number is:

[0031]

[0032] Where, N h Δt is the total number of time steps, H is the rolling update interval, and H is the sliding window size.

[0033] Optionally, the BiLSTM module updates its parameters using a backpropagation algorithm after a cleaning event. Its input is the H heat transfer efficiency sequences accumulated after cleaning, and its output is the predicted value for the next M time steps.

[0034] Secondly, this application provides a multi-cycle energy efficiency prediction device for heat transfer processes that integrates prior knowledge, performing the method described above, including:

[0035] The dataset partitioning module is used to divide the heat transfer efficiency dataset into a cumulative dataset and a sliding window dataset based on the prior knowledge that heat transfer efficiency is affected by fouling and fouling formation.

[0036] The model building module is used to build multi-cycle parallel cascaded ConBiLSTM models, including CNN and BiLSTM modules;

[0037] The condition acquisition module is used to acquire preset event trigger conditions, including prediction error trigger conditions and cleaning operation trigger conditions. When any condition is met, the model update is initiated.

[0038] The determination module is used to dynamically update the accumulated dataset and the sliding window dataset through an adaptive online learning mechanism when the preset event triggering conditions are met.

[0039] The prediction module is used to pause the prediction of the BiLSTM module after the cleaning event and make the initial prediction only through the CNN module until the sliding window dataset has accumulated at least H new samples.

[0040] The output module is used to retrain the CNN module and the BiLSTM module based on the updated dataset, and combine the outputs of the two to generate the final predicted heat transfer efficiency value.

[0041] The adjustment module is used to optimize the maintenance scheduling strategy of the heat transfer system based on the prediction results.

[0042] Optionally, the dataset partitioning module uses GPU acceleration to achieve parallel computation of convolutional and bidirectional long short-term memory networks.

[0043] Thirdly, this application provides a computer device, the device comprising: a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method described above.

[0044] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the method described above.

[0045] In summary, this application includes the following beneficial technical effects:

[0046] This application, based on prior knowledge of the impact of fouling on heat transfer efficiency and fouling formation, divides the heat transfer efficiency dataset into a cumulative dataset and a sliding window dataset. A multi-period parallel cascaded ConBiLSTM model is constructed, including a CNN module and a BiLSTM module. Preset event triggering conditions are obtained. When these conditions are met, the cumulative dataset and the sliding window dataset are dynamically updated through an adaptive online learning mechanism. Initial predictions are made using the CNN module. The CNN and BiLSTM modules are retrained based on the updated dataset, and their outputs are combined to generate the final predicted heat transfer efficiency value. The maintenance and scheduling strategy of the heat transfer system is optimized based on the prediction results. This achieves the technical effect of accurately predicting heat transfer efficiency. Attached Figure Description

[0047] Figure 1This is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiments of this application.

[0048] Figure 2 This is a flowchart illustrating the first embodiment of the multi-cycle energy efficiency prediction method for heat transfer processes that integrates prior knowledge, as described in this application.

[0049] Figure 3 This is a graph showing the changes in heat transfer efficiency over multiple cycles during actual industrial operation of this application.

[0050] Figure 4 This is a diagram of the multi-cycle parallel CNN-BiLSTM framework of this application.

[0051] Figure 5 This is a structural block diagram of the first embodiment of the multi-cycle energy efficiency prediction device for heat transfer processes that integrates prior knowledge, as described in this application. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] Reference Figure 1 , Figure 1 This is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiments of this application.

[0054] like Figure 1 As shown, the computer device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0055] Those skilled in the art will understand that Figure 1The structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0056] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and an event-triggered multi-cycle heat transfer efficiency prediction program for the heat transfer process.

[0057] exist Figure 1 In the computer device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in this application can be set in the computer device. The computer device calls the event-triggered heat transfer process multi-cycle heat transfer efficiency prediction program stored in the memory 1005 through the processor 1001, and executes the event-triggered heat transfer process multi-cycle heat transfer efficiency prediction method provided in the embodiment of this application.

[0058] This application provides a method for predicting the multi-cycle heat transfer efficiency of an event-triggered heat transfer process, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the event-triggered heat transfer process multi-cycle heat transfer efficiency prediction method of this application.

[0059] In this embodiment, the multi-cycle energy efficiency prediction method for heat transfer processes that integrates prior knowledge includes the following steps:

[0060] Step S10: Based on the prior knowledge that heat transfer efficiency is affected by fouling and fouling formation, the heat transfer efficiency dataset is divided into a cumulative dataset and a sliding window dataset.

[0061] The cumulative dataset is used to model spatial nonlinear features, and the sliding window dataset is used to capture short-term time dynamics.

[0062] It should be noted that the change in heat transfer efficiency due to scaling is a complex and dynamic process, exhibiting significant nonlinear behavior and controlled by multiple interacting factors. In the original Kern-Seaton model of scaling, the scaling process is described as consisting of two parts: deposition and descaling, as shown in the following formula.

[0063]

[0064] Among them, the deposition rate (denoted as ) The removal rate (denoted as ) mainly depends on variables such as the surface temperature of the solid-liquid interface and the chemical composition of the deposited material. The scaling rate is influenced by properties such as the viscosity of the scale layer and the fluid flow rate. These factors interact in a complex way, determining the overall scaling rate and its impact on heat transfer efficiency.

[0065] Heat transfer efficiency decreases with increasing scale buildup, exhibiting a clear negative correlation, as shown in the following formula.

[0066]

[0067] Where U(t) is the heat transfer efficiency and U0 is the heat transfer coefficient at the initial moment. Scale accumulation is the most direct and critical factor leading to a decrease in heat transfer efficiency, and heat transfer efficiency can be characterized through the calculation of scale buildup.

[0068] The correlation of scaling rate was considered, and seven key factors were identified. First, density plays a crucial role in scaling formation, affecting the deposition rate of scaling material on the heat transfer surface; surface temperature directly affects solubility, thus influencing scaling; fluid temperature has a complex effect on scaling resistance; the hydraulic diameter of the fluid channel affects scaling by influencing fluid velocity and Reynolds number, both of which are crucial in determining the scaling threshold; fluid velocity affects the degree of scaling by generating shear forces on the material adhering to the heat transfer surface; oxygen content in the fluid accelerates scaling by promoting corrosion and surface roughening; finally, scaling evolves over time, and its accumulation typically increases with time. On the other hand, the accumulation of scaling over time is a key characteristic. Based on the above formulas, a progressive model of scaling is widely used in the control and optimization studies of heat transfer processes, as shown below.

[0069]

[0070] In the model, R f Indicates scaling resistance. It represents its asymptotic value, and τ is a time constant. τ and τ are typically set as constants based on actual industrial processes, while time serves as the primary independent variable, directly affecting the degree of scale accumulation. The importance of this widely accepted model is recognized in this research because it captures the gradual and time-dependent nature of scale accumulation.

[0071] In existing heat transfer efficiency prediction models, the multi-periodic nature is often overlooked. In reality, heat transfer efficiency is influenced by cleaning and maintenance decisions, making it a periodic process, thus resulting in multi-periodic heat transfer efficiency. From an industrial process perspective, periodic cleaning of heat transfer equipment is typically aimed at improving heat transfer efficiency.

[0072] like Figure 3As shown, the entire forecast horizon can be defined as the overall production cycle, where cleaning events divide the forecast timeframe into multiple sub-cycles. Within each sub-cycle, scale gradually accumulates over time, and heat transfer efficiency decreases accordingly.

[0073] It is important to note that the scaling rate is influenced by a variety of factors, which lead to variations in process parameters across different cycles. This makes scaling behavior dynamic and context-dependent, potentially differing from one cycle to the next. Furthermore, while cleaning operations remove accumulated scale from the heat exchanger surface, restoring the scale level to a lower state, the cleaning operation itself does not reset the underlying factors influencing scale accumulation. Instead, it provides a periodic reset of scale values, causing the scaling process to restart and be subject to the same factors. By incorporating multi-cycle factors into the heat transfer efficiency prediction model, this study emphasizes the dynamic nature of scale accumulation and removal, which can lead to more accurate predictions and more efficient maintenance scheduling, ultimately improving the energy efficiency of heat transfer systems and reducing operating costs.

[0074] It should be noted that in the step of dividing the scaling dataset into a cumulative dataset and a sliding window dataset based on prior knowledge of the scaling process, the cumulative dataset X... (s) Through formula

[0075]

[0076] Aggregate historical data, where X ( i) represents the process variable of the i-th stage, and s represents the s-th stage.

[0077] In practical implementation, heat transfer efficiency in industrial systems exhibits significant spatiotemporal complexity due to the nonlinear relationship between inputs and outputs (spatial characteristics) and the temporal evolution trend affected by cleaning operations (temporal characteristics). An adaptive online learning mechanism addresses these complexities by dividing the dataset into spatial and temporal components, enabling independent modeling of long-term spatial relationships and short-term temporal dynamics. Heat transfer efficiency is determined by two key features: spatial nonlinearity and temporal dynamics. On one hand, the nonlinear relationship between operating parameters (such as flow rate and temperature) and fouling resistance requires robust spatial modeling. On the other hand, heat transfer efficiency gradually decreases over time but is periodically reset by cleaning operations, resulting in abrupt discontinuities in the temporal trend. To address these different characteristics, the heat transfer efficiency dataset is divided into two parts: a cumulative dataset that aggregates all historical samples to capture spatial nonlinearity, and a sliding window dataset that is dynamically updated to reflect recent heat transfer efficiency dynamics, ensuring adaptation to short-term temporal variations.

[0078] In practical implementation, the cumulative dataset used for spatial learning is characterized by its continuous growth, providing a comprehensive understanding of spatial features. The sliding window dataset used for temporal learning is characterized by its dynamic updating with the addition of new data, while simultaneously eliminating outdated trends.

[0079] The adaptive mechanism ensures that the time dataset Y (s) Reflects real-time heat transfer efficiency while maintaining the spatial dataset X (s) As a continuously growing record. When updating the time dataset, when the event triggering condition Γ is met. k At this point, the temporal dataset is reset to ensure that predictions are based on the latest heat transfer efficiency. Specifically, the sliding window needs to collect H new samples before it can continue making predictions. Regarding the stability of the spatial dataset, the cumulative dataset X... (s) Unaffected by cleaning and resetting, it maintains a holistic understanding of the input-output relationship.

[0080] Step S20: Construct a multi-cycle parallel cascaded ConBiLSTM model, including a CNN module and a BiLSTM module.

[0081] The CNN module is used to extract the spatial dependencies of process variables, and the BiLSTM module is used to model the bidirectional temporal correlation of heat transfer efficiency.

[0082] In its implementation, this embodiment combines a CNN with a bidirectional BiLSTM network. The CNN module captures the nonlinear spatial relationships between key process variables affecting heat transfer efficiency, which are modeled based on prior knowledge of heat transfer efficiency. Simultaneously, the BiLSTM module enhances the model's ability to capture bidirectional temporal correlations and handle multi-cycle heat transfer efficiency by utilizing historical offline measured heat transfer efficiency data, especially in cases of periodic resets due to cleaning events. The heat transfer efficiency prediction framework is as follows: Figure 4 As shown. For each time step t, the input to the CNN is a feature vector.

[0083]

[0084] Each element corresponds to a specific process variable. The convolutional layer uses multiple convolutional kernels to extract spatial patterns from the seven process variables. Each convolutional kernel... With a width h (where 1 ≤ h ≤ 7), the number of adjacent variables it considers simultaneously is defined. In stage s, the convolution operation of the k-th convolution kernel is:

[0085]

[0086] in, It is the activation value of the k-th convolutional kernel at position i. The input vector X (s) The (i+j)th variable in the equation. Here, b(k) is the weight of the j-th element in the k-th convolutional kernel, b(k) is the bias term, and σ(·) is the activation function. The convolutional layers independently extract the spatial dependencies between variables at each stage. After the convolution operation, the feature maps can be dimensionality-reduced using pooling layers to improve robustness. Then, the pooled features are flattened and passed through fully connected layers to learn the nonlinear relationship between spatial features and heat transfer efficiency.

[0087]

[0088] Where P(s) is the pooling feature map at stage s, and W fc and b fc These are the weight matrix and bias of the fully connected layer, respectively. This is the heat transfer efficiency predicted for stage s.

[0089] The CNN runs continuously, generating predictions throughout the process, including estimated heat transfer efficiency values ​​regardless of whether a cleaning event occurs. As shown below.

[0090]

[0091] The BiLSTM model is used to explore the time-varying pattern of heat transfer efficiency, represented as a sequence.

[0092]

[0093] in, Let represent the true heat transfer efficiency at time step i, and I be the total number of historical time steps used for prediction. BiLSTM processes the input sequence in both forward and backward directions, effectively capturing time dependencies.

[0094]

[0095] The output hidden state at each time step t is obtained by concatenating the forward and reverse hidden states.

[0096]

[0097] The forward and backward propagation of BiLSTM are defined as follows:

[0098] The final heat transfer efficiency prediction combines the outputs of the CNN and BiLSTM modules:

[0099]

[0100] Here, α is a weighting factor used to control the contribution of each module. If due to the cleaning operation (u clean =1) If BiLSTM is paused, the final output depends only on the CNN:

[0101]

[0102] Furthermore, in the proposed multi-cycle parallel cascaded ConBiLSTM framework, both the CNN and BiLSTM modules are designed to adapt to changes in heat transfer efficiency, especially after the event-triggered condition Γ. The CNN is continuously trained to ensure that heat transfer efficiency can still be estimated even when condition Γ occurs and immediately thereafter.

[0103]

[0104] BiLSTM will only resume predictions after collecting at least H new heat transfer efficiency measurements to ensure it has enough data to accurately capture the dynamics after cleaning.

[0105]

[0106] Among them, Y t-H:t It is a sequence of the last H heat transfer efficiency values ​​obtained after cleaning; This represents the predicted heat transfer efficiency for the next M time steps. The BiLSTM model will be retrained when H new data points are obtained.

[0107]

[0108] in, This is the updated parameter set of the BiLSTM after cleaning. This retraining step ensures that the BiLSTM model adapts to the new heat transfer efficiency introduced by the cleaning event.

[0109] It should be noted that the convolution kernel width of the CNN module is 1 to 7, used to simultaneously extract spatial features of adjacent process variables. In the s-th stage, the convolution operation is as follows:

[0110]

[0111] in, It is the activation value of the k-th convolutional kernel at position i. The input vector X (s) In the (i+j)th variable, h represents the width. It is the weight of the j-th element in the k-th convolutional kernel, b (k) σ is the bias term, and σ(·) is the activation function.

[0112] Step S30: Obtain preset event trigger conditions, including prediction error trigger conditions and cleaning operation trigger conditions. Start model update when any condition is met.

[0113] In practical implementation, to improve the efficiency and adaptability of the prediction framework, this embodiment adopts an event-triggered learning strategy. This mechanism ensures that the model is updated only when a significant change in heat transfer efficiency is detected, thereby minimizing unnecessary communication and computational overhead. Two key conditions for triggering model updates are defined. The first is prediction error triggering; when the prediction error within the historical window exceeds a threshold δ1, an update is initiated.

[0114]

[0115] in, It is the prediction error, μ j is the weight of each error term, and h is the size of the historical window.

[0116] The second condition is that a cleaning operation is triggered; when a cleaning operation (t) is detected... k =t clean An update is triggered when ( ).

[0117]

[0118] The final event trigger condition ensures that an update is triggered when there is a high prediction error, a significant change in heat transfer efficiency, or a cleaning event.

[0119]

[0120] Based on the aforementioned concepts of adaptive online learning and event-triggered learning, a hybrid online prediction framework is proposed: a multi-period parallel cascaded ConBiLSTM based on prior knowledge, which extends the CNN-LSTM architecture by combining BiLSTM. Existing research has shown that CNN-LSTM models can effectively capture spatial and temporal features, where CNN is used to extract nonlinear spatial patterns from key process variables, while LSTM is used to model the temporal dependence of heat transfer efficiency. However, these models may face challenges in capturing complex periodic behaviors, especially those caused by cleaning events.

[0121] It should be noted that the step of obtaining the preset event triggering conditions includes prediction error triggering conditions and cleaning operation triggering conditions. The step of starting model update when either condition is met includes:

[0122] Define a prediction error trigger condition: when the prediction error within the historical window exceeds a threshold δ1, an update is initiated.

[0123]

[0124] in, It is the prediction error, μ j is the weight of each error term, and h is the size of the historical window;

[0125] Define the triggering condition for the cleaning operation. When a cleaning operation (t) is detected... k =t clean The update is triggered when )

[0126]

[0127] The preset event triggering conditions are expressed as follows:

[0128]

[0129] Step S40: When it is determined that the preset event triggering conditions are met, the accumulated dataset and the sliding window dataset are dynamically updated through an adaptive online learning mechanism.

[0130] It should be noted that the step of dynamically updating the accumulated dataset and the sliding window dataset through the adaptive online learning mechanism includes:

[0131] After determining that the preset event triggering condition is met. k At that time, the sliding window dataset Y is reset through an adaptive online learning mechanism. (s) The sliding window needs to collect H new samples before it can continue making predictions.

[0132]

[0133] Cumulative dataset X (s) Unaffected by cleaning and resetting, the time dataset Y generated in heat transfer efficiency (s) The formula for calculating the total number is:

[0134]

[0135] Where, N h Δt is the total number of time steps, H is the rolling update interval, and H is the sliding window size.

[0136] Step S50: After the cleaning event, pause the prediction of the BiLSTM module and perform initial prediction only through the CNN module until the sliding window dataset accumulates at least H new samples.

[0137] It should be noted that the BiLSTM module updates its parameters through a backpropagation algorithm after a cleaning event. Its input is the H heat transfer efficiency sequences accumulated after cleaning, and its output is the predicted value for the next M time steps.

[0138] Step S60: Retrain the CNN module and BiLSTM module based on the updated dataset, and combine the outputs of both to generate the final predicted heat transfer efficiency value;

[0139] Step S70: Optimize the maintenance scheduling strategy of the heat transfer system based on the prediction results.

[0140] It should be noted that this embodiment proposes a multi-cycle energy efficiency prediction method for heat transfer processes that integrates prior knowledge, including the following key steps:

[0141] 1. Based on the prior knowledge that heat transfer efficiency is affected by fouling, the dataset is divided into a cumulative dataset and a sliding window dataset.

[0142] 2. Construct a multi-cycle parallel cascaded ConBiLSTM model, which includes a CNN module and a BiLSTM module.

[0143] 3. Set two event trigger conditions: prediction error and cleaning operation.

[0144] 4. When the triggering conditions are met, the dataset is dynamically updated through an adaptive online learning mechanism.

[0145] 5. After cleaning, pause BiLSTM and use CNN for preliminary predictions until enough new samples have been accumulated.

[0146] 6. Retrain the CNN and BiLSTM, and combine the outputs to generate the final prediction.

[0147] This embodiment achieves accurate prediction of multi-cycle changes in heat transfer efficiency by integrating prior knowledge and deep learning models.

[0148] 1. The dataset partitioning utilizes prior knowledge that fouling affects heat transfer efficiency, the cumulative dataset captures spatial nonlinear characteristics, and the sliding window dataset reflects short-term time dynamics.

[0149] 2. The CNN module extracts the spatial dependencies of process variables, and the BiLSTM module models the bidirectional temporal correlation of heat transfer efficiency, together capturing complex nonlinear and temporal features.

[0150] 3. Event triggering conditions ensure timely model updates to adapt to sudden changes in heat transfer efficiency.

[0151] 4. The adaptive online learning mechanism dynamically updates the dataset, enabling the model to adapt to the features of each cycle.

[0152] 5. After cleaning, BiLSTM is paused and only CNN is used for prediction, which avoids the interference of data mutation caused by cleaning on time series prediction.

[0153] 6. By retraining the two modules and combining their outputs, both long-term spatial features are preserved, and new temporal dynamics are captured.

[0154] In practical implementation, the application of this method is illustrated using an industrial heat exchanger as an example:

[0155] 1. Data Preparation: Collect historical operating data of the heat exchanger, including process variables such as flow rate and temperature, as well as heat transfer efficiency data. Based on the prior knowledge that fouling affects heat transfer efficiency, the data is divided into a cumulative dataset (containing all historical data) and a sliding window dataset (containing only data from the most recent period).

[0156] 2. Model Construction: ConBiLSTM model is constructed. The CNN module uses convolutional kernels with widths of 1-7 to extract the spatial relationships between variables such as flow rate and temperature. The BiLSTM module uses a bidirectional LSTM network to capture the temporal dependence of heat transfer efficiency.

[0157] 3. Set trigger conditions: Define prediction error trigger conditions (such as prediction error exceeding 5%) and cleaning operation trigger conditions.

[0158] 4. Model Prediction and Update:

[0159] - During normal operation, the model continuously predicts heat transfer efficiency.

[0160] - A model update is triggered when the prediction error exceeds 5% or a cleaning operation is detected.

[0161] - During updates, dynamically adjust the sliding window dataset to retain the latest data.

[0162] - If triggered by a cleaning operation, pause BiLSTM prediction and use only CNN for initial prediction.

[0163] - After collecting enough new data (e.g., 100 new samples), retrain the CNN and BiLSTM modules.

[0164] - Combine the outputs of the two modules to generate the final predicted heat transfer efficiency.

[0165] 5. Maintenance and Optimization: Develop a cleaning plan for the heat exchanger based on the forecast results. For example, if the forecast indicates that the heat transfer efficiency will drop below the threshold within a week, schedule cleaning operations in advance.

[0166] In this way, the method can accurately predict the changes in heat transfer efficiency of heat exchangers during multiple fouling-cleaning cycles, providing strong support for equipment maintenance and energy efficiency optimization.

[0167] This design enables the model to simultaneously handle the spatial nonlinearity and temporal dynamics of heat transfer efficiency, especially adapting well to periodic changes caused by cleaning, thus achieving accurate prediction of multi-cycle heat transfer efficiency.

[0168] It should be noted that, to address the complex fouling-cleaning characteristics in multi-fluid heat exchange systems, this embodiment presents an adaptive multi-model ensemble prediction method based on fluid characteristic classification. The core of this method is to establish an independent ConBiLSTM model for each major fluid type and achieve adaptive ensemble of the models through a dynamic weight adjustment mechanism.

[0169] The specific implementation steps are as follows:

[0170] 1. Fluid Classification: Based on the main fluid type in the system (such as crude oil, natural gas, chemical solvents, etc.), historical data are classified according to fluid type.

[0171] 2. Multiple Model Construction: Independent ConBiLSTM models are constructed for each major fluid type. The structure of each model is the same as the original scheme, but it is trained using historical data of the corresponding fluid type.

[0172] 3. Enhanced Feature Extraction: Add fluid-related features, such as viscosity, density, heat capacity and other physicochemical parameters, to the CNN module to better capture the scaling characteristics of different fluids.

[0173] 4. Dynamic weight adjustment mechanism: Design a weight allocator to dynamically adjust the prediction weights of each fluid-specific model based on the proportion and operating status of various fluids in the current system.

[0174] 5. Integrated Prediction: The prediction results of each fluid-specific model are weighted and averaged according to dynamically allocated weights to obtain the final predicted value of heat transfer efficiency.

[0175] 6. Adaptive Update: The event-triggered conditions and adaptive online learning mechanism of the original scheme are retained, but updates are performed separately for each fluid-specific model. When the prediction error of a certain fluid exceeds a threshold or a cleaning operation occurs, only the corresponding fluid-specific model is updated.

[0176] 7. Post-cleaning processing optimization: After the cleaning operation, depending on the specific type of fluid cleaned, the BiLSTM module of the corresponding fluid-specific model is selectively paused, and only the CNN module is used for preliminary prediction until enough new samples are accumulated.

[0177] This adaptive multi-model ensemble method based on fluid property classification can more accurately capture the unique behavior of different fluids during the scaling-cleaning cycle. By establishing a dedicated model for each fluid and dynamically adjusting the prediction weights according to real-time operating conditions, this method can effectively adapt to the complex scaling-cleaning characteristics of multi-fluid heat exchange systems, improving overall prediction accuracy. Simultaneously, by retaining the adaptive update mechanism from the original scheme, this method can continuously optimize the performance of each fluid-specific model, ensuring the accuracy and reliability of long-term predictions.

[0178] In this embodiment, a multi-cycle heat transfer efficiency prediction model based on prior knowledge is proposed. This framework dynamically updates its predictions by continuously improving the CNN model as new heat transfer efficiency data is acquired. Initially, the CNN component is used to predict heat transfer efficiency values, relying on historical data to model the current heat transfer efficiency decline process in real time. After each cleaning event, the model is updated to reflect the post-cleaning state and restart the heat transfer efficiency decline process. At the beginning of each cycle, only the CNN is used for prediction. However, as additional test data accumulates within the cycle, the BiLSTM component is activated, enabling the model to capture the temporal dynamics of heat transfer efficiency and the unique periodic changes of each operational phase. This adaptive learning mechanism allows the framework to gradually improve its prediction accuracy with each update, ensuring that it closely tracks the observed heat transfer efficiency patterns and reduces prediction errors over time. This model effectively combines prior knowledge of heat transfer efficiency and cleaning events with a dynamic learning method, considering the spatial and temporal characteristics of the heat transfer process. This embodiment designs an adaptive event-triggered learning framework to address the multi-cycle characteristics of heat transfer efficiency, especially the periodic reset problem caused by cleaning operations. The framework incorporates an adaptive mechanism that pauses predictions during a cleaning event and updates the model with new data after cleaning. This ensures the model responds to changes in operating conditions and consistently provides reliable predictions. Experimental results validate the effectiveness of the proposed framework, demonstrating a significant improvement in prediction accuracy.

[0179] This embodiment, based on prior knowledge of heat transfer efficiency, divides the heat transfer efficiency dataset into a cumulative dataset and a sliding window dataset, constructs a multi-cycle parallel cascaded ConBiLSTM model, including a CNN module and a BiLSTM module, obtains preset event trigger conditions, and dynamically updates the cumulative dataset and sliding window dataset through an adaptive online learning mechanism when the preset event trigger conditions are met; initial prediction is performed through the CNN module; the CNN module and BiLSTM module are retrained based on the updated dataset, and the final heat transfer efficiency prediction value is generated by combining their outputs; the maintenance and scheduling strategy of the heat transfer system is optimized based on the prediction results. This achieves the technical effect of accurately predicting heat transfer efficiency.

[0180] Furthermore, embodiments of this application also propose a computer-readable storage medium storing a program for predicting the multi-cycle heat transfer efficiency of an event-triggered heat transfer process. When the program for predicting the multi-cycle heat transfer efficiency of an event-triggered heat transfer process is executed by a processor, it implements the steps of the method for predicting the multi-cycle heat transfer efficiency of an event-triggered heat transfer process as described above.

[0181] Reference Figure 5 , Figure 5 This is a structural block diagram of the first embodiment of the multi-cycle energy efficiency prediction device for heat transfer processes that integrates prior knowledge, as described in this application.

[0182] like Figure 5 As shown in the embodiments of this application, the multi-cycle energy efficiency prediction device for heat transfer processes that integrates prior knowledge includes:

[0183] The dataset partitioning module 10 is used to partition the heat transfer efficiency dataset into a cumulative dataset and a sliding window dataset based on prior knowledge of heat transfer efficiency.

[0184] Model building module 20 is used to build a multi-cycle parallel cascaded ConBiLSTM model, including a CNN module and a BiLSTM module;

[0185] The condition acquisition module 30 is used to acquire preset event triggering conditions, including prediction error triggering conditions and cleaning operation triggering conditions, and to start model update when any condition is met.

[0186] The determination module 40 is used to dynamically update the accumulated dataset and the sliding window dataset through an adaptive online learning mechanism when the preset event triggering conditions are met.

[0187] Prediction module 50 is used to pause the prediction of the BiLSTM module after the cleaning event and make initial predictions only through the CNN module until the sliding window dataset accumulates at least H new samples.

[0188] Output module 60 is used to retrain the CNN module and BiLSTM module based on the updated dataset, and combine the outputs of the two to generate the final predicted value of heat transfer efficiency.

[0189] Adjustment module 70 is used to optimize the maintenance scheduling strategy of the heat transfer system based on the prediction results.

[0190] It should be noted that the dataset partitioning module uses GPU acceleration to achieve parallel computation of convolutional and bidirectional long short-term memory networks.

[0191] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solution of this application. In specific applications, those skilled in the art can make settings as needed, and this application does not impose any restrictions on this.

[0192] This embodiment, based on prior knowledge of the impact of fouling on heat transfer efficiency and the formation of fouling, divides the heat transfer efficiency dataset into a cumulative dataset and a sliding window dataset. A multi-period parallel cascaded ConBiLSTM model is constructed, including a CNN module and a BiLSTM module. Preset event triggering conditions are obtained. When these conditions are met, the cumulative dataset and the sliding window dataset are dynamically updated through an adaptive online learning mechanism. Initial predictions are made using the CNN module. The CNN and BiLSTM modules are retrained based on the updated dataset, and their outputs are combined to generate the final predicted heat transfer efficiency value. The maintenance and scheduling strategy of the heat transfer system is optimized based on the prediction results. This achieves the technical effect of accurately predicting heat transfer efficiency.

[0193] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0194] In addition, for technical details not described in detail in this embodiment, please refer to the method for predicting the multi-cycle heat transfer efficiency of the event-triggered heat transfer process provided in any embodiment of this application, which will not be repeated here.

[0195] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0196] The sequence numbers of the embodiments in this application are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for multi-period energy efficiency prediction of heat transfer process with priori knowledge fusion, characterized in that, The method comprises the steps of: dividing a heat transfer efficiency dataset into a cumulative dataset and a sliding window dataset based on prior knowledge of the influence of fouling on heat transfer efficiency and the formation of fouling, wherein the cumulative dataset is used to model spatial nonlinear characteristics, and the sliding window dataset is used to capture short-term time dynamics; building a multi-cycle parallel cascaded ConBiLSTM model, which comprises a CNN module and a BiLSTM module, wherein the CNN module is used to extract the spatial dependence of process variables, and the BiLSTM module is used to model the bidirectional time correlation of heat transfer efficiency; obtaining preset event trigger conditions, including a prediction error trigger condition and a cleaning operation trigger condition, and starting model updating when any condition is met; when it is determined that the preset event trigger condition is met, dynamically updating the cumulative dataset and the sliding window dataset through an adaptive online learning mechanism; after a cleaning event, pausing the prediction of the BiLSTM module, performing initial prediction only through the CNN module, and accumulating at least H new samples in the sliding window dataset; retraining the CNN module and the BiLSTM module according to the updated dataset, and generating a final heat transfer efficiency prediction value by combining the outputs of the two modules; optimizing the maintenance scheduling strategy of the heat transfer system based on the prediction result.

2. The method of claim 1, wherein, The step of dividing the heat transfer efficiency dataset into a cumulative dataset X (s) By the formula aggregated historical data, wherein X (i) is the process variable for the i-th stage, and s denotes the s-th stage.

3. The method of claim 1, wherein, The CNN module has a convolution kernel width of 1 to 7, which is used to simultaneously extract the spatial features of adjacent process variables, and in the s stage, the convolution operation is: where, is the activation value of the kth convolution kernel at position i, is the input vector X (s) the (i + j)th variable in the input vector X, h denotes the width, is the weight of the jth element in the kth convolution kernel, b (k) is the bias term, and σ(·) is the activation function.

4. The method of claim 1, wherein, The step of obtaining preset event trigger conditions, including a prediction error trigger condition and a cleaning operation trigger condition, and starting model updating when any condition is met, comprises: defining a prediction error trigger condition, wherein when the prediction error in the historical window exceeds a threshold value δ1, the updating is started: wherein, is the prediction error, μ j is the weight of each error term, h is the history window size; Define a cleaning operation trigger condition, which triggers an update when a cleaning operation (t k = t clean ) is detected, The preset event trigger condition is represented as:

5. The method of claim 1, wherein, The step of dynamically updating the cumulative dataset and the sliding window dataset through an adaptive online learning mechanism comprises: When it is determined that the preset event triggering condition Γ k is met, the sliding window dataset Y (s) is reset through an adaptive online learning mechanism, and the sliding window needs to collect H new samples before the prediction can continue. Cumulative data set X (s) Time data set Y generated in the heat transfer period, not affected by the purge reset (s) The total number is calculated by the formula: where N h is the total number of time steps, At is the rolling update interval, and H is the sliding window size.

6. The method of claim 1, wherein, The BiLSTM module updates parameters through a back propagation algorithm after a cleaning event, and the input is H heat transfer efficiency sequences accumulated after cleaning, and the output is prediction values for M future time steps.

7. A heat transfer process multi-period energy efficiency prediction device fusing prior knowledge, characterized by, The method comprises the steps of: a dataset division module for dividing a heat transfer efficiency dataset into a cumulative dataset and a sliding window dataset based on prior knowledge of a multi-cycle energy efficiency prediction of a heat transfer process; a model building module for building a multi-cycle parallel cascaded ConBiLSTM model, which comprises a CNN module and a BiLSTM module; a condition acquisition module for obtaining preset event trigger conditions, including a prediction error trigger condition and a cleaning operation trigger condition, and starting model updating when any condition is met; a determination module for dynamically updating the cumulative dataset and the sliding window dataset through an adaptive online learning mechanism when it is determined that the preset event trigger condition is met; a prediction module for pausing the prediction of the BiLSTM module after a cleaning event, performing initial prediction only through the CNN module, and accumulating at least H new samples in the sliding window dataset. The output module is configured to retrain the CNN module and the BiLSTM module according to the updated data set, and combine outputs of the two modules to generate a final heat transfer efficiency prediction value. The adjustment module is configured to optimize a maintenance scheduling strategy of the heat transfer system based on the prediction result.

8. The apparatus of claim 7, wherein, The data set division module realizes parallel computation of the convolution and the bidirectional long short-term memory network through GPU acceleration.

9. A computer device, comprising: The device comprises a memory and a processor, and the processor executes the method according to any one of claims 1-6 when running computer instructions stored in the memory.

10. A computer-readable storage medium, characterized in that, The instructions, when running on a computer, cause the computer to execute the method according to any one of claims 1-6.

Images