A deep learning-based heat pump intelligent cleaning method and system

Abstract

The application discloses a kind of based on deep learning's heat pump intelligent cleaning method and system, including step one: periodically collecting heat pump operating state data, and according to the order of collection time constitutes heat pump operating dataset;Step two: according to heat pump operating dataset, determine whether heat pump needs cleaning;Step three: if it is determined that heat pump needs cleaning, then heat pump operating dataset is input into improved N-BEATS network to carry out cleaning representation prediction;Step four: determine target cleaning site and fouling grade;Step five: generate stage-by-stage cleaning scheme, and execute cleaning operation to target cleaning site;Step six: in the process of executing cleaning operation, whether it is determined to switch current main cleaning stage in real time;Step seven: confirm cleaning result after cleaning operation ends.The application improves the pertinence of heat pump intelligent cleaning by improved N-BEATS network and stage-by-stage cleaning control.

Description

Technical Field

[0001] This invention relates to the field of heat pump cleaning technology, and in particular to a heat pump intelligent cleaning method and system based on deep learning. Background Technology

[0002] With the widespread application of heat pump equipment in building heating, industrial waste heat recovery, agricultural drying, and centralized hot water supply, intelligent cleaning and control technologies targeting fouling accumulation, flow channel blockage, and performance degradation in heat pump heat exchange components are receiving increasing attention. Existing heat pump cleaning methods mainly rely on manual inspections, periodic shutdowns for maintenance, or simple trigger controls based on single operating parameter thresholds for cleaning determination. However, these methods commonly suffer from the following problems in practical applications: During long-term operation, the accumulation and blockage evolution of fouling in the evaporator, condenser, and heat exchange channels of a heat pump exhibits significant gradualness, coupling, and time-varying characteristics. Existing technologies typically rely solely on single data points such as temperature, pressure, or flow rate for judgment, making it difficult to accurately distinguish between fouling accumulation, channel blockage, and performance changes caused by short-term operational fluctuations. This leads to high misjudgment rates and insufficient reliability of cleaning decisions. Heat pump operating status data exhibits continuous temporal variations, and complex correlations exist between temperature differences, pressure differences, flow rates, and power at different sampling times. Existing judgment methods based on empirical rules or simple statistical analysis are insufficient to fully characterize the fouling development process, failing to effectively predict the degree of fouling, the rate of fouling development, and the expected duration of continued operation. This results in delayed or premature cleaning selection, impacting heat pump operating efficiency and maintenance costs. For different heat exchange components and different causes of fouling, traditional cleaning control methods often employ a uniform cleaning process or fixed cleaning stage switching logic. They lack the ability to generate phased cleaning plans and dynamically control stage switching based on cleaning attribution results and fouling levels, leading to insufficient cleaning targeting, wasted cleaning resources, and unstable cleaning results.

[0003] Therefore, how to provide a heat pump intelligent cleaning method and system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based intelligent cleaning method and system for heat pumps. This invention fully utilizes an improved N-BEATS network and details the heat pump cleaning determination, cleaning attribution analysis, cleaning timing prediction, target cleaning location identification, phased cleaning control, and cleaning result confirmation. It possesses advantages such as high accuracy in cleaning determination, precise timing of cleaning, highly targeted cleaning process, and high cleaning efficiency.

[0005] A deep learning-based intelligent cleaning method for heat pumps according to an embodiment of the present invention includes the following steps: Step 1: Within the current operating cycle, periodically collect heat pump operating status data and construct a heat pump operating dataset according to the order of collection time; Step 2: Based on the heat pump operation dataset, calculate the cleaning decision coefficient sequence of the heat pump in the current operation cycle, and determine whether the heat pump needs to be cleaned based on the cleaning decision coefficient sequence, and output the cleaning attribution result; Step 3: If it is determined that the heat pump needs cleaning, the heat pump operation dataset of the current operating cycle is input into the improved N-BEATS network to predict the degree of fouling, the rate of fouling development, and the expected duration of continued operation of the heat pump, generating a cleaning timing representation vector. The improved N-BEATS network includes an input mapping module, several cascaded dual-path residual prediction blocks, and an output accumulation module. The dual-path residual prediction blocks introduce a dual-path parallel prediction structure and a residual back-substitution correction structure. Step 4: Based on the cleaning attribution results and the cleaning timing characterization vector, perform zonal analysis on different heat exchange parts of the heat pump to determine the target cleaning parts and fouling levels; Step 5: Based on the cleaning attribution results and the level of fouling, generate a phased cleaning plan, and perform cleaning operations on the target cleaning areas according to the phased cleaning plan; Step Six: During the cleaning operation, determine in real time whether to switch to the current main cleaning stage; Step 7: After the cleaning operation is completed, acquire the heat pump operating status data and confirm the cleaning results based on the changes in operating status before and after cleaning.

[0006] Optionally, the heat pump operating status data includes evaporator inlet temperature, evaporator outlet temperature, condenser inlet temperature, condenser outlet temperature, evaporator inlet pressure, evaporator outlet pressure, condenser inlet pressure, condenser outlet pressure, heat exchange medium circulation flow rate, compressor operating power, outdoor ambient temperature, outdoor ambient humidity, user-side return water temperature, and user-side circulation flow rate.

[0007] Optionally, step two specifically includes: Calculate the difference between the evaporator inlet temperature and the evaporator outlet temperature, and the difference between the condenser inlet temperature and the condenser outlet temperature, respectively, and add them together to obtain a comprehensive temperature difference characterization quantity. Calculate the difference between the evaporator inlet pressure and the evaporator outlet pressure, and the difference between the condenser inlet pressure and the condenser outlet pressure, respectively, and add them together to obtain a comprehensive pressure difference characterization quantity. Calculate the average of the heat exchange medium circulation flow rate and the user-side circulation flow rate to obtain the flow correction amount; Calculate the product of the comprehensive pressure difference characterization quantity and the compressor operating power to obtain the resistance work characterization quantity; calculate the product of the comprehensive temperature difference characterization quantity and the comprehensive pressure difference characterization quantity, and add it to the constant 1 to obtain the heat transfer correction quantity; Calculate the ratio of the resistance work characteristic quantity to the heat transfer correction quantity to obtain the cleaning determination coefficient; Within the current operating cycle, the cleaning judgment coefficients are arranged in the order of collection time to form a cleaning judgment coefficient sequence, and the cycle cleaning judgment mean and cycle cleaning judgment increase are calculated based on the cleaning judgment coefficient sequence. The change in cleaning judgment is calculated based on the cleaning judgment coefficient of adjacent collection times, and the average change is generated based on all the changes in cleaning judgment. If the average value of the periodic cleaning judgment is greater than the cleaning judgment coefficient corresponding to the first collection time, and the increase in the periodic cleaning judgment is greater than 0, then it is determined that the heat pump has a cleaning abnormality in the current operating cycle. For an operating cycle with cleaning anomalies, if the change in each cleaning judgment is greater than or equal to 0, and the change in each cleaning judgment is less than or equal to the average change, then the cleaning attribution result is determined to be dirt accumulation, and the heat pump needs to be cleaned. If the change in cleaning determination is greater than the average change, and all cleaning determination coefficients after the next acquisition time corresponding to the current change in cleaning determination remain continuous and do not decrease, then the cleaning attribution result is that the flow channel is blocked, and the heat pump needs to be cleaned. If the change in cleaning determination is greater than the average change, and all cleaning determination coefficients after the next collection time corresponding to the current change in cleaning determination decrease, then the cleaning attribution result is determined to be short-term operational fluctuation, and the heat pump does not need to be cleaned.

[0008] Optionally, step three specifically includes: The heat pump operation dataset is represented as an input feature vector sequence according to the acquisition time dimension and the feature dimension; In the input mapping module, the input feature vector sequence is mapped using a linear mapping layer and the GELU activation function to generate an initial hidden feature vector sequence. Each dual-path residual prediction block includes a shared feature extraction unit, a trend prediction path, a mutation prediction path, a path fusion unit, a multi-level residual back substitution unit, and a prediction output unit; The input features of the first dual-path residual prediction block are the initial hidden feature vector sequence. Each dual-path residual prediction block generates a corrected residual vector sequence and a block-level prediction vector by performing dual-path parallel modeling, path fusion, multi-level residual back substitution and prediction output on the input features. The next dual-path residual prediction block receives the corrected residual vector sequence generated by the previous dual-path residual prediction block as input features; In the output accumulation module, the block-level prediction vectors of all dual-path residual prediction blocks are accumulated to obtain the cleaning timing characterization vector. The first component of the cleaning timing characterization vector represents the fouling degree characterization value, the second component represents the fouling development speed, and the third component represents the expected duration of continued operation.

[0009] Optionally, the dual-path residual prediction block introduces a dual-path parallel prediction structure and a residual back-substitution correction structure, specifically including: In the shared feature extraction unit, the input features are extracted through an alternating structure consisting of two linear mapping layers and two GELU activation functions to obtain a sequence of shared hidden feature vectors. The trend prediction path employs a dual-scale pooling reconstruction structure to generate a sequence of trend explanation feature vectors, specifically: The shared hidden feature vector sequence is subjected to average pooling in two time windows of different scales to obtain the first smooth feature vector sequence and the second smooth feature vector sequence. The first and second smoothed feature vector sequences are restored to the original acquisition time dimension by linear interpolation, and then the features are concatenated to obtain the trend concatenated feature vector sequence. The trend concatenation feature vector sequence is passed through a linear mapping layer and the GELU activation function to generate a trend interpretation feature vector sequence. The mutation prediction path employs a differential convolutional response structure to generate a sequence of mutation explanation feature vectors, specifically: The shared hidden feature vector sequence is used to obtain the differential feature vector sequence through feature difference operation between adjacent acquisition times, and the first feature vector of the differential feature vector sequence is the zero vector; The difference feature vector sequence is mapped to the local mutation response through a one-dimensional convolutional layer, a GELU activation function, and another one-dimensional convolutional layer to obtain the mutation explanation feature vector sequence. In the path fusion unit, the trend explanation feature vector sequence and the mutation explanation feature vector sequence are combined through feature concatenation, linear mapping layer and GELU activation function to generate joint explanation feature vector sequence; The multi-level residual back substitution unit generates a corrected residual vector sequence by performing residual interpretation, residual back substitution, and residual correction on the joint interpretation feature vector sequence, specifically as follows: The joint explanatory feature vector sequence is passed through a linear mapping layer to generate a first-level explanatory vector sequence; the difference between the joint explanatory feature vector sequence and the first-level explanatory vector sequence is calculated to obtain the first-level residual vector sequence; The joint interpretation feature vector sequence and the first-level residual vector sequence are used to generate a back-substitution feature vector sequence through feature concatenation, a linear mapping layer, and the GELU activation function. The back-substituted feature vector sequence is subjected to residual correction through a linear mapping layer to generate a corrected intermediate vector sequence. Calculate the difference between the first-level residual vector sequence and the corrected intermediate vector sequence to generate the corrected residual vector sequence; In the prediction output unit, the joint explanation feature vector sequence and the corrected residual vector sequence are aggregated to generate a block-level fused feature vector; the block-level fused feature vector is then passed through a linear mapping layer, a GELU activation function, and another linear mapping layer to generate a block-level prediction vector.

[0010] Optionally, step four specifically includes: The zoning analysis includes the analysis of the cleaning areas of the evaporator, the condenser, and the heat exchange channel. Based on the heat pump operating status data, the evaporator temperature difference, evaporator pressure difference, condenser temperature difference, and condenser pressure difference are calculated, and the flow correction amount is obtained. If the cleaning cause is dirt accumulation, calculate the dirt determination value for the evaporator cleaning area and the condenser cleaning area based on the evaporator temperature difference, evaporator pressure difference, condenser temperature difference and condenser pressure difference, respectively. The cleaning area with a higher dirt determination value is selected as the target cleaning area, and the dirt level coefficient is obtained by calculating the product of the dirt clogging degree characterization value, the dirt clogging development rate, and the constant 1. If the cleaning cause is channel blockage, calculate the blockage judgment value for the evaporator cleaning part, condenser cleaning part, and heat exchange channel cleaning part based on the evaporator temperature difference, evaporator pressure difference, condenser temperature difference, condenser pressure difference, and flow correction amount. The cleaning area with the larger clogging judgment value is selected as the target cleaning area, and the clogging level coefficient is obtained by calculating the ratio of the clogging degree characterization value to the sum of the expected continuing operating time and the constant 1. Based on the fouling level coefficient and the clogging level coefficient, threshold comparisons are performed to determine the fouling level of fouling accumulation and flow channel blockage; the fouling level is divided into Level 1 fouling, Level 2 fouling and Level 3 fouling according to the cleaning attribution results.

[0011] Optionally, step five specifically includes: If the cleaning result indicates dirt buildup, the phased cleaning plan is as follows: When the fouling level is Level 1 fouling, the process includes a pre-rinse stage, a main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level 2 fouling, the process includes a pre-rinse stage, a main cleaning stage, an extended main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level three, the process includes a pre-rinse stage, a dual-circulation main cleaning stage, a rinsing stage, and a recovery stage. If the cleaning result indicates flow channel blockage, the phased cleaning plan is as follows: When the fouling level is Level 1 fouling, the process includes a pre-rinse stage, an enhanced flushing stage, a rinsing stage, and a recovery stage. When the fouling level is level 2 fouling, the process includes a pre-rinse stage, an enhanced flushing stage, a main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level three, the process includes a pre-rinse stage, a two-stage enhanced flushing stage, a main cleaning stage, a rinsing stage, and a recovery stage.

[0012] Optionally, step six specifically includes: During the main cleaning phase, the temperature difference recovery value and pressure difference recovery value of the target cleaning area are periodically acquired; the main cleaning phase includes the main cleaning phase, the main cleaning extension phase, the dual-cycle main cleaning phase, the enhanced rinsing phase, and the dual-stage enhanced rinsing phase; The change in temperature difference recovery is calculated based on the temperature difference recovery value of adjacent sampling times, and the change in pressure difference recovery is calculated based on the pressure difference recovery value of adjacent sampling times. If the temperature difference recovery change is less than or equal to the temperature difference switching threshold, and the pressure difference recovery change is less than or equal to the pressure difference switching threshold, the current main cleaning stage is determined to have met the switching conditions, and the process is switched to the next main cleaning stage or rinsing stage. If the switching conditions are not met, the current main cleaning phase will continue to be executed. The temperature difference switching threshold is the average value of all temperature difference recovery changes in the latter half of the current main cleaning stage; the pressure difference switching threshold is the average value of all pressure difference recovery changes in the latter half of the current main cleaning stage.

[0013] Optionally, the cleaning result includes cleaning completed and cleaning abnormal. If the cleaning result is cleaning abnormal, the main cleaning stage is re-executed based on the cleaning attribution result.

[0014] A deep learning-based intelligent heat pump cleaning system according to an embodiment of the present invention includes: The operation data acquisition module is used to periodically collect heat pump operation status data within the current operation cycle and construct a heat pump operation dataset according to the acquisition time sequence. The cleaning determination module is used to calculate the cleaning determination coefficient sequence of the heat pump in the current operating cycle based on the heat pump operation dataset, and determine whether the heat pump needs to be cleaned based on the cleaning determination coefficient sequence, and output the cleaning attribution result. The cleaning timing prediction module is used to input the heat pump operation dataset of the current operating cycle into the improved N-BEATS network when it is determined that the heat pump needs to be cleaned, and to predict the fouling degree characterization value, fouling development speed and expected continued operation time of the heat pump, and generate a cleaning timing characterization vector. The cleaning part analysis module is used to perform zonal analysis on different heat exchange parts of the heat pump based on the cleaning attribution results and the cleaning timing characterization vector, and to determine the target cleaning part and the fouling level. The cleaning solution generation module is used to generate phased cleaning solutions based on cleaning attribution results and fouling levels. The cleaning execution module is used to perform cleaning operations on the target cleaning area according to the phased cleaning plan; The stage switching control module is used to determine in real time whether to switch the current main cleaning stage during the cleaning operation. The cleaning result confirmation module is used to acquire heat pump operating status data after the cleaning operation is completed, and to confirm the cleaning result based on the changes in operating status before and after cleaning.

[0015] The beneficial effects of this invention are: This invention periodically collects heat pump operating status data to construct a heat pump operating dataset. Based on a cleaning determination coefficient sequence, it differentiates and determines fouling accumulation, flow channel blockage, and short-term operational fluctuations, thereby accurately identifying whether the heat pump needs cleaning during its operation and accurately determining the specific reasons for the cleaning requirement. In the cleaning timing prediction stage, the heat pump operating dataset of the current operating cycle is input into an improved N-BEATS network with a dual-path parallel prediction structure and a residual back-substitution correction structure. This network jointly predicts the fouling degree characterization value, fouling development speed, and expected continued operating time, improving the ability to characterize the heat pump fouling evolution process and cleaning timing. In the cleaning execution stage, combining the cleaning attribution results, cleaning timing characterization vector, and fouling level, the cleaning areas of the evaporator, condenser, and heat exchange channels are analyzed in zones, generating a phased cleaning plan. Simultaneously, during the main cleaning stages, the phase switching conditions are determined in real time based on the temperature difference recovery value and pressure difference recovery value, thereby improving the targeting of cleaning area identification, the dynamics of cleaning process control, and the reliability of cleaning result confirmation. Therefore, this invention is of great significance for improving the accuracy of heat pump cleaning determination, the precision of cleaning timing prediction, the effect of phased cleaning control, and the overall operation and maintenance efficiency. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1This is a schematic diagram of a heat pump intelligent cleaning method and system based on deep learning proposed in this invention; Figure 2 This invention presents a deep learning-based intelligent heat pump cleaning method and a flowchart of heat pump cleaning determination and cleaning cause attribution in the system. Figure 3 This is a flowchart of an improved N-BEATS network structure in a heat pump intelligent cleaning method and system based on deep learning proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A deep learning-based intelligent cleaning method for heat pumps includes the following steps: Step 1: Within the current operating cycle, periodically collect heat pump operating status data and construct a heat pump operating dataset according to the order of collection time; Step 2: Based on the heat pump operation dataset, calculate the cleaning decision coefficient sequence of the heat pump in the current operation cycle, and determine whether the heat pump needs to be cleaned based on the cleaning decision coefficient sequence, and output the cleaning attribution result; Step 3: If it is determined that the heat pump needs cleaning, the heat pump operation dataset of the current operating cycle is input into the improved N-BEATS network to predict the degree of fouling, the rate of fouling development, and the expected duration of continued operation of the heat pump, generating a cleaning timing representation vector. The improved N-BEATS network includes an input mapping module, several cascaded dual-path residual prediction blocks, and an output accumulation module. The dual-path residual prediction blocks introduce a dual-path parallel prediction structure and a residual back-substitution correction structure. Step 4: Based on the cleaning attribution results and the cleaning timing characterization vector, perform zonal analysis on different heat exchange parts of the heat pump to determine the target cleaning parts and fouling levels; Step 5: Based on the cleaning attribution results and the level of fouling, generate a phased cleaning plan, and perform cleaning operations on the target cleaning areas according to the phased cleaning plan; Step Six: During the cleaning operation, determine in real time whether to switch to the current main cleaning stage; Step 7: After the cleaning operation is completed, acquire the heat pump operating status data and confirm the cleaning results based on the changes in operating status before and after cleaning.

[0019] In this embodiment, the heat pump operating status data includes evaporator inlet temperature, evaporator outlet temperature, condenser inlet temperature, condenser outlet temperature, evaporator inlet pressure, evaporator outlet pressure, condenser inlet pressure, condenser outlet pressure, heat exchange medium circulation flow rate, compressor operating power, outdoor ambient temperature, outdoor ambient humidity, user-side return water temperature, and user-side circulation flow rate.

[0020] In this embodiment, step two specifically includes: Calculate the difference between the evaporator inlet temperature and the evaporator outlet temperature, and the difference between the condenser inlet temperature and the condenser outlet temperature, respectively, and add them together to obtain a comprehensive temperature difference characterization quantity. Calculate the difference between the evaporator inlet pressure and the evaporator outlet pressure, and the difference between the condenser inlet pressure and the condenser outlet pressure, respectively, and add them together to obtain a comprehensive pressure difference characterization quantity. Calculate the average of the heat exchange medium circulation flow rate and the user-side circulation flow rate to obtain the flow correction amount; Calculate the product of the comprehensive pressure difference characterization quantity and the compressor operating power to obtain the resistance work characterization quantity; calculate the product of the comprehensive temperature difference characterization quantity and the comprehensive pressure difference characterization quantity, and add it to the constant 1 to obtain the heat transfer correction quantity; Calculate the ratio of the resistance work characteristic quantity to the heat transfer correction quantity to obtain the cleaning determination coefficient; In this invention, the cleaning determination coefficient is used to characterize the intensity of the cleaning demand of the heat pump as reflected by the combined effects of heat exchange performance degradation, flow resistance changes, and operating power changes at the current sampling time. The overall heat exchange state on both sides of the evaporator and condenser is reflected by a comprehensive temperature difference, the flow resistance state within the heat exchange channel is reflected by a comprehensive pressure difference, and the energy cost borne by the heat pump to maintain its current operating state is characterized by the compressor operating power. This unifies the degree of heat exchange deterioration and the degree of flow channel obstruction into a comparable determination quantity. When there is fouling or flow channel blockage inside the heat pump, the heat exchange efficiency decreases, the pressure loss increases, and the compressor load increases, thus increasing the cleaning determination coefficient. Therefore, the cleaning determination coefficient can comprehensively characterize the degree of performance degradation of the heat pump caused by contamination and blockage.

[0021] Within the current operating cycle, the cleaning judgment coefficients are arranged in the order of collection time to form a cleaning judgment coefficient sequence, and the cycle cleaning judgment mean and cycle cleaning judgment increase are calculated based on the cleaning judgment coefficient sequence; where the cycle cleaning judgment mean is the mean of the cleaning judgment coefficient sequence; the cycle cleaning judgment increase is the difference between the cleaning judgment coefficient corresponding to the last collection time and the cleaning judgment coefficient corresponding to the first collection time within the current operating cycle. The change in cleaning judgment is calculated based on the cleaning judgment coefficient of adjacent collection times, and the average change is generated based on all the changes in cleaning judgment. If the average value of the periodic cleaning judgment is greater than the cleaning judgment coefficient corresponding to the first collection time, and the increase in the periodic cleaning judgment is greater than 0, then it is determined that the heat pump has a cleaning abnormality in the current operating cycle. For an operating cycle with cleaning anomalies, if the change in each cleaning judgment is greater than or equal to 0, and the change in each cleaning judgment is less than or equal to the average change, then the cleaning attribution result is determined to be dirt accumulation, and the heat pump needs to be cleaned. If the change in cleaning determination is greater than the average change, and all cleaning determination coefficients after the next acquisition time corresponding to the current change in cleaning determination remain continuous and do not decrease, then the cleaning attribution result is that the flow channel is blocked, and the heat pump needs to be cleaned. If the change in cleaning determination is greater than the average change, and all cleaning determination coefficients after the next collection time corresponding to the current change in cleaning determination decrease, then the cleaning attribution result is determined to be short-term operational fluctuation, and the heat pump does not need to be cleaned.

[0022] In this embodiment, step three specifically includes: The heat pump operation dataset is represented as an input feature vector sequence according to the acquisition time dimension and the feature dimension; In the input mapping module, the input feature vector sequence is mapped using a linear mapping layer and the GELU activation function to generate an initial hidden feature vector sequence. Each dual-path residual prediction block includes a shared feature extraction unit, a trend prediction path, a mutation prediction path, a path fusion unit, a multi-level residual back substitution unit, and a prediction output unit; The input features of the first dual-path residual prediction block are the initial hidden feature vector sequence. Each dual-path residual prediction block generates a corrected residual vector sequence and a block-level prediction vector by performing dual-path parallel modeling, path fusion, multi-level residual back substitution and prediction output on the input features. The next dual-path residual prediction block receives the corrected residual vector sequence generated by the previous dual-path residual prediction block as input features; In the output accumulation module, the block-level prediction vectors of all dual-path residual prediction blocks are accumulated to obtain the cleaning timing characterization vector. The first component of the cleaning timing characterization vector represents the fouling degree characterization value, the second component represents the fouling development speed, and the third component represents the expected duration of continued operation.

[0023] In this embodiment, the dual-path residual prediction block introduces a dual-path parallel prediction structure and a residual back-substitution correction structure, specifically including: In the shared feature extraction unit, the input features are extracted through an alternating structure consisting of two linear mapping layers and two GELU activation functions to obtain a sequence of shared hidden feature vectors. The trend prediction path employs a dual-scale pooling reconstruction structure to generate a sequence of trend-explanatory feature vectors, specifically: The shared hidden feature vector sequence is subjected to average pooling in two time windows of different scales to obtain the first smooth feature vector sequence and the second smooth feature vector sequence. The first and second smoothed feature vector sequences are restored to the original acquisition time dimension by linear interpolation, and then the features are concatenated to obtain the trend concatenated feature vector sequence. The trend concatenation feature vector sequence is passed through a linear mapping layer and the GELU activation function to generate a trend interpretation feature vector sequence. The mutation prediction path employs a differential convolutional response structure to generate a sequence of mutation explanatory feature vectors, specifically: The shared hidden feature vector sequence is used to obtain the differential feature vector sequence through feature difference operation between adjacent acquisition times, wherein the first feature vector of the differential feature vector sequence is the zero vector; The difference feature vector sequence is mapped to the local mutation response through a one-dimensional convolutional layer, a GELU activation function, and another one-dimensional convolutional layer to obtain the mutation explanation feature vector sequence. In the path fusion unit, the trend explanation feature vector sequence and the mutation explanation feature vector sequence are combined through feature concatenation, linear mapping layer and GELU activation function to generate joint explanation feature vector sequence; The multi-level residual back substitution unit generates a corrected residual vector sequence by performing residual interpretation, residual back substitution, and residual correction on the joint interpretation feature vector sequence, specifically as follows: The joint explanatory feature vector sequence is passed through a linear mapping layer to generate a first-level explanatory vector sequence; the difference between the joint explanatory feature vector sequence and the first-level explanatory vector sequence is calculated to obtain the first-level residual vector sequence; The joint interpretation feature vector sequence and the first-level residual vector sequence are used to generate a back-substitution feature vector sequence through feature concatenation, a linear mapping layer, and the GELU activation function. The back-substituted feature vector sequence is subjected to residual correction through a linear mapping layer to generate a corrected intermediate vector sequence. Calculate the difference between the first-level residual vector sequence and the corrected intermediate vector sequence to generate the corrected residual vector sequence; In the prediction output unit, the joint explanation feature vector sequence and the corrected residual vector sequence are aggregated to generate a block-level fused feature vector; the block-level fused feature vector is then passed through a linear mapping layer, a GELU activation function, and another linear mapping layer to generate a block-level prediction vector.

[0024] In this invention, the improved N-BEATS network retains the basic framework and core processing logic compared to the standard N-BEATS network. The standard N-BEATS network includes an input mapping part, multiple cascaded prediction blocks, and an output aggregation part. The input time series first enters the network for feature representation, and then passes through multiple prediction blocks layer by layer. Each prediction block interprets the current input sequence and outputs the corresponding prediction result. The residual information obtained after processing by the previous prediction block is passed to the next prediction block, enabling subsequent prediction blocks to continue modeling the parts that have not been fully interpreted by the previous blocks. Finally, the prediction results output by all prediction blocks are accumulated to obtain the final prediction result.

[0025] The improved N-BEATS network reconstructs the internal structure of a single prediction block. In the standard N-BEATS network, a single prediction block uses a single-path feature extraction method, where input features are continuously mapped to directly form the interpretation and prediction results, resulting in a relatively simple information flow within the block. The improved N-BEATS network replaces this with a dual-path residual prediction block, introducing a dual-path parallel prediction structure and a residual back-substitution correction structure within the block. Specifically, after the shared feature extraction unit, the network no longer extracts temporal features along a single path but splits them into a trend prediction path and a mutation prediction path. The trend prediction path uses a dual-scale pooling reconstruction structure to extract smooth changes and long-term evolution information from different time scales; the mutation prediction path uses a differential convolution response structure to extract local fluctuations and abrupt changes between adjacent acquisition times. Subsequently, in the path fusion unit, the trend interpretation feature vector sequence and the mutation interpretation feature vector sequence are fused to form a joint interpretation feature vector sequence. In addition, the standard N-BEATS network completes one interpretation and one residual update within a block, while the improved N-BEATS network further adds multi-level residual back-substitution units. Through first-level interpretation, residual back-substitution and residual correction, a corrected residual vector sequence is formed, so that the residual propagation within the block is no longer a single deduction, but has the characteristics of step-by-step correction.

[0026] The improved N-BEATS network extracts long-term trend information through trend prediction paths and local fluctuation information through abrupt change prediction paths. This enables the network to characterize the gradual evolution of heat pump fouling and sudden changes in operating status, thereby improving its ability to identify fouling severity and development speed. Simultaneously, multi-level residual back-substitution units can progressively interpret and correct the joint interpretation feature vector sequence, reducing information loss caused by single residual interpretation and improving the accuracy of residual sequence correction. This makes the input features received by subsequent dual-path residual prediction blocks more targeted. Based on these improvements, the improved N-BEATS network makes the predictions of the expected continued operating time and cleaning timing representation vectors more stable and accurate, which is more conducive to subsequent target cleaning site analysis, fouling level determination, and generation of phased cleaning plans.

[0027] In this embodiment, step four specifically includes: The zoning analysis includes the analysis of the cleaning areas of the evaporator, condenser, and heat exchange channels. Based on the heat pump operating status data, the evaporator temperature difference, evaporator pressure difference, condenser temperature difference, and condenser pressure difference are calculated, and the flow correction amount is obtained. If the cleaning cause is determined to be dirt accumulation, the dirt determination values ​​for the evaporator cleaning area and the condenser cleaning area are calculated based on the evaporator temperature difference, evaporator pressure difference, condenser temperature difference, and condenser pressure difference, respectively. The dirt determination value for the evaporator cleaning area is the ratio of the sum of the evaporator temperature difference and evaporator pressure difference to a constant 1; the dirt determination value for the condenser cleaning area is the ratio of the sum of the condenser temperature difference and condenser pressure difference to a constant 1. The cleaning area with a higher dirt determination value is selected as the target cleaning area, and the dirt level coefficient is obtained by calculating the product of the dirt clogging degree characterization value, the dirt clogging development rate, and the constant 1. If the cleaning cause is channel blockage, calculate the blockage determination values ​​for the evaporator cleaning section, condenser cleaning section, and heat exchange channel cleaning section based on the evaporator temperature difference, evaporator pressure difference, condenser temperature difference, condenser pressure difference, and flow correction. Specifically, the blockage determination value for the evaporator cleaning section is the ratio of the sum of the evaporator pressure difference and the evaporator temperature difference to a constant 1; the blockage determination value for the condenser cleaning section is the ratio of the sum of the condenser pressure difference and the condenser temperature difference to a constant 1; and the blockage determination value for the heat exchange channel cleaning section is the reciprocal of the sum of the flow correction and a constant 1. The cleaning area with the larger clogging judgment value is selected as the target cleaning area, and the clogging level coefficient is obtained by calculating the ratio of the clogging degree characterization value to the sum of the expected continuing operating time and the constant 1. Based on the fouling level coefficient and the clogging level coefficient, threshold comparisons are performed to determine the fouling level of fouling accumulation and flow channel blockage; among them, the fouling level is divided into Level 1 fouling, Level 2 fouling and Level 3 fouling according to the cleaning attribution results.

[0028] In this invention, for the purpose of addressing dirt accumulation, based on a sample set of dirt level coefficients from historical operating cycles, the samples are sorted in ascending order. The dirt level coefficient corresponding to one-third of the positions is taken as the first dirt level threshold, and the dirt level coefficient corresponding to two-thirds of the positions is taken as the second dirt level threshold. Specifically, a first-level blockage is defined as a dirt level coefficient less than the first dirt level threshold; a second-level blockage is defined as a dirt level coefficient less than the second dirt level threshold but greater than or equal to the first dirt level threshold; and a third-level blockage is defined as a dirt level coefficient greater than the second dirt level threshold. For flow channel blockage, based on the blockage level coefficient sample set in the historical operating cycle, the samples are sorted in ascending order. The blockage level coefficient corresponding to one-third of the position is taken as the first blockage level threshold, and the blockage level coefficient corresponding to two-thirds of the position is taken as the second blockage level threshold. Specifically, level 1 blockage is defined as a blockage level coefficient less than the first blockage level threshold; level 2 blockage is defined as a blockage level coefficient less than the second blockage level threshold and greater than or equal to the first blockage level threshold; and level 3 blockage is defined as a blockage level coefficient greater than the second blockage level threshold.

[0029] In this embodiment, step five specifically includes: If the cleaning result indicates dirt buildup, the phased cleaning plan is as follows: When the fouling level is Level 1 fouling, the process includes a pre-rinse stage, a main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level 2 fouling, the process includes a pre-rinse stage, a main cleaning stage, an extended main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level three, the process includes a pre-rinse stage, a dual-circulation main cleaning stage, a rinsing stage, and a recovery stage. If the cleaning result indicates flow channel blockage, the phased cleaning plan is as follows: When the fouling level is Level 1 fouling, the process includes a pre-rinse stage, an enhanced flushing stage, a rinsing stage, and a recovery stage. When the fouling level is level 2 fouling, the process includes a pre-rinse stage, an enhanced flushing stage, a main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level three, the process includes a pre-rinse stage, a two-stage enhanced flushing stage, a main cleaning stage, a rinsing stage, and a recovery stage.

[0030] In this embodiment, step six specifically includes: During the main cleaning phase, the temperature difference recovery value and pressure difference recovery value of the target cleaning area are periodically acquired; the main cleaning phase includes the main cleaning phase, the extended main cleaning phase, the dual-cycle main cleaning phase, the enhanced rinsing phase, and the dual-stage enhanced rinsing phase. The change in temperature difference recovery is calculated based on the temperature difference recovery value of adjacent sampling times, and the change in pressure difference recovery is calculated based on the pressure difference recovery value of adjacent sampling times. If the temperature difference recovery change is less than or equal to the temperature difference switching threshold, and the pressure difference recovery change is less than or equal to the pressure difference switching threshold, the current main cleaning stage is determined to have met the switching conditions, and the process is switched to the next main cleaning stage or rinsing stage. If the switching conditions are not met, the current main cleaning phase will continue to be executed. Among them, the temperature difference switching threshold is the average value of all temperature difference recovery changes in the second half of the current main cleaning stage; the pressure difference switching threshold is the average value of all pressure difference recovery changes in the second half of the current main cleaning stage.

[0031] In this embodiment, the cleaning result includes cleaning completed and cleaning abnormal. If the cleaning result is cleaning abnormal, the main cleaning stage is re-executed based on the cleaning attribution result.

[0032] A deep learning-based intelligent heat pump cleaning system includes: The operation data acquisition module is used to periodically collect heat pump operation status data within the current operation cycle and construct a heat pump operation dataset according to the acquisition time sequence. The cleaning determination module is used to calculate the cleaning determination coefficient sequence of the heat pump in the current operating cycle based on the heat pump operation dataset, and determine whether the heat pump needs to be cleaned based on the cleaning determination coefficient sequence, and output the cleaning attribution result. The cleaning timing prediction module is used to input the heat pump operation dataset of the current operating cycle into the improved N-BEATS network when it is determined that the heat pump needs to be cleaned, and to predict the fouling degree characterization value, fouling development speed and expected continued operation time of the heat pump, and generate a cleaning timing characterization vector. The cleaning part analysis module is used to perform zonal analysis on different heat exchange parts of the heat pump based on the cleaning attribution results and the cleaning timing characterization vector, and to determine the target cleaning part and the fouling level. The cleaning solution generation module is used to generate phased cleaning solutions based on cleaning attribution results and fouling levels. The cleaning execution module is used to perform cleaning operations on the target cleaning area according to the phased cleaning plan; The stage switching control module is used to determine in real time whether to switch the current main cleaning stage during the cleaning operation. The cleaning result confirmation module is used to acquire heat pump operating status data after the cleaning operation is completed, and to confirm the cleaning result based on the changes in operating status before and after cleaning.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to the cleaning and maintenance of air source heat pump units in an energy station of a commercial complex. This energy station is responsible for the mall's winter heating, transitional season domestic hot water supply, and some areas with constant temperature loads. A total of six air source heat pump units with a rated heating capacity of 320kW were configured on-site. Unit No. 2, which had a long continuous operating time, was prone to scaling on its heat exchanger surface, and experienced significant fluctuations in user-side load, was selected as the implementation target. This unit had been running continuously for 14 months, during which only one routine shutdown flushing was performed. Maintenance personnel found that the unit exhibited slower outlet water temperature rise, increased compressor operating power, decreased heat exchange medium circulation flow, and increased fluctuations in user-side return water temperature. Traditional methods mainly rely on manual inspection experience and fixed-cycle cleaning, often only scheduling shutdowns after significant scale accumulation. This leads to problems such as inaccurate timing of cleaning, difficulty in distinguishing short-term operational fluctuations from actual fouling, and unstable recovery after cleaning.

[0034] In the implementation of this invention, temperature sensors, pressure sensors, flow sensors, and electrical parameter acquisition units deployed on the heat pump unit periodically collect data on the evaporator inlet temperature, evaporator outlet temperature, condenser inlet temperature, condenser outlet temperature, evaporator inlet pressure, evaporator outlet pressure, condenser inlet pressure, condenser outlet pressure, heat exchange medium circulation flow rate, compressor operating power, outdoor ambient temperature, outdoor ambient humidity, user-side return water temperature, and user-side circulation flow rate. These data are then collected in chronological order to form a heat pump operation dataset. The on-site acquisition cycle is set to 5 minutes, forming an operation cycle data packet every 24 hours. After continuously recording 21 days of operation data, the system calculates the cleaning judgment coefficient sequence. The results show that, starting on day 15, the average value of the periodic cleaning judgment for heat pump unit No. 2 consistently exceeds the cleaning judgment coefficient corresponding to the first sampling point, and the increase in periodic cleaning judgment is continuously positive. By day 18, some changes in cleaning judgment are significantly greater than the average change, but the subsequent cleaning judgment coefficient does not decrease rapidly but remains continuously. The system attributes the cleaning to flow channel blockage and determines that the unit needs cleaning. Subsequently, the heat pump operation dataset for this cycle was input into an improved N-BEATS network to predict the fouling degree characterization value, fouling development rate, and expected continued operating time, resulting in a cleaning timing characterization vector. The fouling degree characterization value was 0.81, the fouling development rate was 0.064 / day, and the expected continued operating time was 3.6 days. Based on these results, the system identified the heat exchange channel cleaning area as the primary target cleaning area, determined the fouling level to be level three, and automatically generated a phased cleaning plan consisting of a pre-rinse stage, a two-stage intensive rinsing stage, a main cleaning stage, a rinsing stage, and a recovery operation stage.

[0035] In practical applications, the system first performs a pre-rinse using a low-flow-rate circulation method to initially remove loose deposits from the heat exchange channels. This is followed by a two-stage enhanced rinsing stage, where alternating forward and reverse rinsing and intermittent pressurized rinsing are applied to the heat exchange channels. Finally, the system enters the main cleaning stage for deep cleaning. During the cleaning process, the system continuously reads the temperature difference recovery value and pressure difference recovery value of the target cleaning area. Based on the average value of all temperature difference recovery changes and pressure difference recovery changes in the latter half of the main cleaning stage, a stage switching threshold is established. When the temperature difference recovery change and pressure difference recovery change are both less than the corresponding threshold, the system automatically switches to the next main cleaning stage or rinsing stage, thus avoiding over-rinsing or under-rinsing problems caused by traditional fixed-duration cleaning. After cleaning, the system again collects heat pump operating status data and confirms the cleaning results based on the changes in operating status before and after cleaning.

[0036] To verify the actual effect of the present invention, under the same load range, similar outdoor ambient temperature conditions, and the same test cycle, the present invention was compared and analyzed with two comparative schemes. Comparative scheme one is a method of manual experience judgment plus fixed cycle cleaning, while comparative scheme two uses a standard N-BEATS network for cleaning timing prediction and cleaning control. The comparison results are shown in Table 1.

[0037] Table 1. Comparison of Application Effects of Different Heat Pump Cleaning Solutions

[0038] As shown in Table 1, the present invention outperforms both Comparative Scheme 1 and Comparative Scheme 2 in all comparative indicators. Compared with Comparative Scheme 1, the cleaning judgment accuracy of the present invention increased from 78.6% to 95.8%, and the false judgment rate decreased from 21.4% to 4.2%, indicating that the present invention can more accurately distinguish between dirt accumulation, flow channel blockage, and short-term operational fluctuations, reducing false judgments and missed judgments. Compared with Comparative Scheme 2, the cleaning judgment accuracy of the present invention further increased by 6.1 percentage points, and the false judgment rate further decreased by 6.1 percentage points, indicating that the improved N-BEATS network used in the present invention has higher accuracy in characterizing heat pump fouling and predicting cleaning timing.

[0039] Meanwhile, the cleaning timing advance of this invention reaches 3.4 days, which is higher than the 0.8 days of comparative scheme one and the 2.3 days of comparative scheme two, indicating that this invention can identify heat pump cleaning needs earlier and arrange cleaning operations in advance. Regarding the cleaning recovery effect, the heating efficiency recovery rate after cleaning reaches 19.6%, the temperature difference recovery range reaches 3.4℃, and the pressure difference reduction reaches 38.9 kPa, all significantly better than the comparative schemes, indicating that this invention can more effectively restore the heat pump heat exchange performance and flow channel unobstructed state. Furthermore, the cleaning time per cycle is shortened to 111 minutes, and the number of fault warnings within 7 days after cleaning is reduced to 0, further demonstrating that this invention can enhance operational stability after cleaning while improving cleaning efficiency.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A heat pump intelligent cleaning method based on deep learning, characterized in that, Includes the following steps: Step 1: Within the current operating cycle, periodically collect heat pump operating status data and construct a heat pump operating dataset according to the order of collection time; Step 2: Based on the heat pump operation dataset, calculate the cleaning decision coefficient sequence of the heat pump in the current operation cycle, and determine whether the heat pump needs to be cleaned based on the cleaning decision coefficient sequence, and output the cleaning attribution result; Step 3: If it is determined that the heat pump needs cleaning, the heat pump operation dataset of the current operating cycle is input into the improved N-BEATS network to predict the degree of fouling, the rate of fouling development, and the expected duration of continued operation of the heat pump, and to generate a cleaning timing representation vector; the improved N-BEATS network includes an input mapping module, several cascaded dual-path residual prediction blocks, and an output accumulation module; The dual-path residual prediction block introduces a dual-path parallel prediction structure and a residual back-substitution correction structure. Step 4: Based on the cleaning attribution results and the cleaning timing characterization vector, perform zonal analysis on different heat exchange parts of the heat pump to determine the target cleaning parts and fouling levels; Step 5: Based on the cleaning attribution results and the level of fouling, generate a phased cleaning plan, and perform cleaning operations on the target cleaning areas according to the phased cleaning plan; Step Six: During the cleaning operation, determine in real time whether to switch to the current main cleaning stage; Step 7: After the cleaning operation is completed, acquire the heat pump operating status data and confirm the cleaning results based on the changes in operating status before and after cleaning.

2. The heat pump intelligent cleaning method based on deep learning according to claim 1, characterized in that, The heat pump operating status data includes evaporator inlet temperature, evaporator outlet temperature, condenser inlet temperature, condenser outlet temperature, evaporator inlet pressure, evaporator outlet pressure, condenser inlet pressure, condenser outlet pressure, heat exchange medium circulation flow rate, compressor operating power, outdoor ambient temperature, outdoor ambient humidity, user-side return water temperature, and user-side circulation flow rate.

3. The heat pump intelligent cleaning method based on deep learning according to claim 1, characterized in that, Step two specifically includes: Calculate the difference between the evaporator inlet temperature and the evaporator outlet temperature, and the difference between the condenser inlet temperature and the condenser outlet temperature, respectively, and add them together to obtain a comprehensive temperature difference characterization quantity. Calculate the difference between the evaporator inlet pressure and the evaporator outlet pressure, and the difference between the condenser inlet pressure and the condenser outlet pressure, respectively, and add them together to obtain a comprehensive pressure difference characterization quantity. Calculate the average of the heat exchange medium circulation flow rate and the user-side circulation flow rate to obtain the flow correction amount; Calculate the product of the comprehensive pressure difference characterization quantity and the compressor operating power to obtain the resistance work characterization quantity; calculate the product of the comprehensive temperature difference characterization quantity and the comprehensive pressure difference characterization quantity, and add it to the constant 1 to obtain the heat transfer correction quantity; Calculate the ratio of the resistance work characteristic quantity to the heat transfer correction quantity to obtain the cleaning determination coefficient; Within the current operating cycle, the cleaning judgment coefficients are arranged in the order of collection time to form a cleaning judgment coefficient sequence, and the cycle cleaning judgment mean and cycle cleaning judgment increase are calculated based on the cleaning judgment coefficient sequence. The change in cleaning judgment is calculated based on the cleaning judgment coefficient of adjacent collection times, and the average change is generated based on all the changes in cleaning judgment. If the average value of the periodic cleaning judgment is greater than the cleaning judgment coefficient corresponding to the first collection time, and the increase in the periodic cleaning judgment is greater than 0, then it is determined that the heat pump has a cleaning abnormality in the current operating cycle. For an operating cycle with cleaning anomalies, if the change in each cleaning judgment is greater than or equal to 0, and the change in each cleaning judgment is less than or equal to the average change, then the cleaning attribution result is determined to be dirt accumulation, and the heat pump needs to be cleaned. If the change in cleaning determination is greater than the average change, and all cleaning determination coefficients after the next acquisition time corresponding to the current change in cleaning determination remain continuous and do not decrease, then the cleaning attribution result is that the flow channel is blocked, and the heat pump needs to be cleaned. If the change in cleaning determination is greater than the average change, and all cleaning determination coefficients after the next collection time corresponding to the current change in cleaning determination decrease, then the cleaning attribution result is determined to be short-term operational fluctuation, and the heat pump does not need to be cleaned.

4. The heat pump intelligent cleaning method based on deep learning according to claim 1, characterized in that, Step three specifically includes: The heat pump operation dataset is represented as an input feature vector sequence according to the acquisition time dimension and the feature dimension; In the input mapping module, the input feature vector sequence is mapped using a linear mapping layer and the GELU activation function to generate an initial hidden feature vector sequence. Each dual-path residual prediction block includes a shared feature extraction unit, a trend prediction path, a mutation prediction path, a path fusion unit, a multi-level residual back substitution unit, and a prediction output unit; The input features of the first dual-path residual prediction block are the initial hidden feature vector sequence. Each dual-path residual prediction block generates a corrected residual vector sequence and a block-level prediction vector by performing dual-path parallel modeling, path fusion, multi-level residual back substitution and prediction output on the input features. The next dual-path residual prediction block receives the corrected residual vector sequence generated by the previous dual-path residual prediction block as input features; In the output accumulation module, the block-level prediction vectors of all dual-path residual prediction blocks are accumulated to obtain the cleaning timing characterization vector. The first component of the cleaning timing characterization vector represents the fouling degree characterization value, the second component represents the fouling development speed, and the third component represents the expected duration of continued operation.

5. The heat pump intelligent cleaning method based on deep learning according to claim 1, characterized in that, The dual-path residual prediction block introduces a dual-path parallel prediction structure and a residual back-substitution correction structure, specifically including: In the shared feature extraction unit, the input features are extracted through an alternating structure consisting of two linear mapping layers and two GELU activation functions to obtain a sequence of shared hidden feature vectors. The trend prediction path employs a dual-scale pooling reconstruction structure to generate a sequence of trend explanation feature vectors, specifically: The shared hidden feature vector sequence is subjected to average pooling in two time windows of different scales to obtain the first smooth feature vector sequence and the second smooth feature vector sequence. The first and second smoothed feature vector sequences are restored to the original acquisition time dimension by linear interpolation, and then the features are concatenated to obtain the trend concatenated feature vector sequence. The trend concatenation feature vector sequence is passed through a linear mapping layer and the GELU activation function to generate a trend interpretation feature vector sequence. The mutation prediction path employs a differential convolutional response structure to generate a sequence of mutation explanation feature vectors, specifically: The shared hidden feature vector sequence is used to obtain the differential feature vector sequence through feature difference operation between adjacent acquisition times, and the first feature vector of the differential feature vector sequence is the zero vector; The difference feature vector sequence is mapped to the local mutation response through a one-dimensional convolutional layer, a GELU activation function, and another one-dimensional convolutional layer to obtain the mutation explanation feature vector sequence. In the path fusion unit, the trend explanation feature vector sequence and the mutation explanation feature vector sequence are combined through feature concatenation, linear mapping layer and GELU activation function to generate joint explanation feature vector sequence; The multi-level residual back substitution unit generates a corrected residual vector sequence by performing residual interpretation, residual back substitution, and residual correction on the joint interpretation feature vector sequence, specifically as follows: The joint explanatory feature vector sequence is passed through a linear mapping layer to generate a first-level explanatory vector sequence; the difference between the joint explanatory feature vector sequence and the first-level explanatory vector sequence is calculated to obtain the first-level residual vector sequence; The joint interpretation feature vector sequence and the first-level residual vector sequence are used to generate a back-substitution feature vector sequence through feature concatenation, a linear mapping layer, and the GELU activation function. The back-substituted feature vector sequence is subjected to residual correction through a linear mapping layer to generate a corrected intermediate vector sequence. Calculate the difference between the first-level residual vector sequence and the corrected intermediate vector sequence to generate the corrected residual vector sequence; In the prediction output unit, the joint explanation feature vector sequence and the corrected residual vector sequence are aggregated to generate a block-level fused feature vector; the block-level fused feature vector is then passed through a linear mapping layer, a GELU activation function, and another linear mapping layer to generate a block-level prediction vector.

6. The heat pump intelligent cleaning method based on deep learning according to claim 1, characterized in that, Step four specifically includes: The zoning analysis includes the analysis of the cleaning areas of the evaporator, condenser, and heat exchange channels; Based on the heat pump operating status data, the evaporator temperature difference, evaporator pressure difference, condenser temperature difference, and condenser pressure difference are calculated, and the flow correction amount is obtained. If the cleaning cause is dirt accumulation, calculate the dirt determination value for the evaporator cleaning area and the condenser cleaning area based on the evaporator temperature difference, evaporator pressure difference, condenser temperature difference and condenser pressure difference, respectively. The cleaning area with a higher dirt determination value is selected as the target cleaning area, and the dirt level coefficient is obtained by calculating the product of the dirt clogging degree characterization value, the dirt clogging development rate, and the constant 1. If the cleaning cause is channel blockage, calculate the blockage judgment value for the evaporator cleaning part, condenser cleaning part, and heat exchange channel cleaning part respectively based on the evaporator temperature difference, evaporator pressure difference, condenser temperature difference, condenser pressure difference, and flow correction amount. The cleaning area with the larger clogging judgment value is selected as the target cleaning area, and the clogging level coefficient is obtained by calculating the ratio of the clogging degree characterization value to the sum of the expected continuing operating time and the constant 1. Based on the fouling level coefficient and the clogging level coefficient, threshold comparisons are performed to determine the fouling level of fouling accumulation and flow channel blockage; the fouling level is divided into Level 1 fouling, Level 2 fouling and Level 3 fouling according to the cleaning attribution results.

7. The heat pump intelligent cleaning method based on deep learning according to claim 1, characterized in that, Step five specifically includes: If the cleaning result indicates dirt buildup, the phased cleaning plan is as follows: When the fouling level is Level 1 fouling, the process includes a pre-rinse stage, a main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level 2 fouling, the process includes a pre-rinse stage, a main cleaning stage, an extended main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level three, the process includes a pre-rinse stage, a dual-circulation main cleaning stage, a rinsing stage, and a recovery stage. If the cleaning result indicates flow channel blockage, the phased cleaning plan is as follows: When the fouling level is Level 1 fouling, the process includes a pre-rinse stage, an enhanced flushing stage, a rinsing stage, and a recovery stage. When the fouling level is level 2 fouling, the process includes a pre-rinse stage, an enhanced flushing stage, a main cleaning stage, a rinsing stage, and a recovery stage. When the fouling level is level three, the process includes a pre-rinse stage, a two-stage enhanced flushing stage, a main cleaning stage, a rinsing stage, and a recovery stage.

8. The heat pump intelligent cleaning method based on deep learning according to claim 1, characterized in that, Step six specifically includes: During the main cleaning phase, the temperature difference recovery value and pressure difference recovery value of the target cleaning area are periodically acquired; the main cleaning phase includes the main cleaning phase, the extended main cleaning phase, the dual-cycle main cleaning phase, the enhanced rinsing phase, and the dual-stage enhanced rinsing phase; The change in temperature difference recovery is calculated based on the temperature difference recovery value of adjacent sampling times, and the change in pressure difference recovery is calculated based on the pressure difference recovery value of adjacent sampling times. If the temperature difference recovery change is less than or equal to the temperature difference switching threshold, and the pressure difference recovery change is less than or equal to the pressure difference switching threshold, the current main cleaning stage is determined to have met the switching conditions, and the process is switched to the next main cleaning stage or rinsing stage. If the switching conditions are not met, the current main cleaning phase will continue to be executed. The temperature difference switching threshold is the average value of all temperature difference recovery changes in the latter half of the current main cleaning stage; the pressure difference switching threshold is the average value of all pressure difference recovery changes in the latter half of the current main cleaning stage.

9. The heat pump intelligent cleaning method based on deep learning according to claim 1, characterized in that, The cleaning results include cleaning completed and cleaning abnormal. If the cleaning result is cleaning abnormal, the main cleaning stage will be re-executed based on the cleaning attribution result.

10. A deep learning-based intelligent heat pump cleaning system, comprising performing the deep learning-based intelligent heat pump cleaning method according to any one of claims 1 to 9, characterized in that, include: The operation data acquisition module is used to periodically collect heat pump operation status data within the current operation cycle and construct a heat pump operation dataset according to the acquisition time sequence. The cleaning determination module is used to calculate the cleaning determination coefficient sequence of the heat pump in the current operating cycle based on the heat pump operation dataset, and determine whether the heat pump needs to be cleaned based on the cleaning determination coefficient sequence, and output the cleaning attribution result. The cleaning timing prediction module is used to input the heat pump operation dataset of the current operating cycle into the improved N-BEATS network when it is determined that the heat pump needs to be cleaned, and to predict the fouling degree characterization value, fouling development speed and expected continued operation time of the heat pump, and generate a cleaning timing characterization vector. The cleaning part analysis module is used to perform zonal analysis on different heat exchange parts of the heat pump based on the cleaning attribution results and the cleaning timing characterization vector, and to determine the target cleaning part and the fouling level. The cleaning solution generation module is used to generate phased cleaning solutions based on cleaning attribution results and fouling levels. The cleaning execution module is used to perform cleaning operations on the target cleaning area according to the phased cleaning plan; The stage switching control module is used to determine in real time whether to switch the current main cleaning stage during the cleaning operation. The cleaning result confirmation module is used to acquire heat pump operating status data after the cleaning operation is completed, and to confirm the cleaning result based on the changes in operating status before and after cleaning.

Images