Air source heat pump electric energy consumption prediction method
By constructing a frost susceptibility index and a virtual frost accumulation, the energy consumption of air source heat pumps is decomposed. A dual-channel prediction structure is used for optimized training, which solves the problem of low accuracy caused by frost and defrosting in the prediction of energy consumption of air source heat pumps and achieves high-precision energy consumption prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- KUNMING HERUI TECH
- Filing Date
- 2026-05-11
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for predicting the power consumption of air source heat pumps have low accuracy in dealing with the additional energy consumption caused by frost and defrosting, especially under high humidity and low temperature conditions where there is a systematic bias. Furthermore, existing methods fail to effectively quantify the gradual increase in compressor power caused by frost growth.
By collecting and preprocessing data, a frost susceptibility index and virtual frost accumulation are constructed, frost intervals are identified and energy consumption is decomposed, and a dual-channel prediction structure is used to deduce the steady-state heating base energy consumption and the defrosting pulse additional energy consumption respectively. The training effect is optimized by combining near-critical oversampling and cross-backtracking verification to achieve accurate prediction of energy consumption.
It significantly improves the accuracy and reliability of predicting the electricity consumption of air source heat pumps, effectively quantifies the additional energy consumption caused by frosting and defrosting, avoids the measurement bias of traditional methods, and enhances the engineering practicality and adaptability of energy consumption prediction.
Smart Images
Figure CN122288043A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of energy consumption monitoring and prediction technology for air source heat pumps, specifically a method for predicting the electrical energy consumption of air source heat pumps. Background Technology
[0002] Air source heat pumps have become an important technical route to replace traditional coal-fired and gas-fired boilers in the field of building heating due to their flexible installation and wide applicability. Accurate prediction of the power consumption of air source heat pumps is of great significance for building energy consumption assessment, power load scheduling, operation strategy optimization and heating cost accounting.
[0003] Existing prediction methods are divided into three categories: physical models, data-driven methods, and hybrid physical and data methods. When dealing with the actual operating energy consumption of air source heat pumps, these three methods share a common weakness: they lack accurate means to quantify the additional energy consumption caused by evaporator frosting and defrosting.
[0004] Under heating conditions, frost reduces the heat pump's heat exchange efficiency, and the triggered reverse-cycle defrosting process generates additional energy consumption. Furthermore, the timing and frequency of defrosting are highly uncertain over time due to the interaction of multiple factors. Existing physical models use fixed empirical correction coefficients to handle the effects of defrosting, which cannot adapt to changes in meteorological conditions. Data-driven methods tend to treat the pulsed power spikes of defrosting as noise smoothing, and tend to underestimate energy consumption under high humidity and low temperature conditions and overestimate energy consumption under dry and low temperature conditions. Hybrid physical and data methods calculate the baseline energy consumption through a physical model and then use a data-driven model to correct the residuals of the physical model. Such methods, to a certain extent, balance physical interpretability and data adaptability, but their model structure is complex and still does not address the aforementioned common weaknesses.
[0005] The additional energy consumption from frost formation and defrosting accounts for 10%–30% of the total energy consumption of a heat pump under typical heating weather conditions, but existing methods have a prediction error of over 50%. At the same time, the gradual increase in compressor power caused by frost growth is also ignored by existing methods, further increasing the cumulative error in energy consumption prediction and leading to a systematic deviation in energy consumption prediction under high humidity and low temperature frost conditions.
[0006] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention
[0007] The purpose of this invention is to solve the problem of low prediction accuracy caused by abnormal fluctuations in energy consumption of air source heat pumps due to frosting and defrosting, and thus proposes a method for predicting the electrical energy consumption of air source heat pumps.
[0008] The objective of this invention can be achieved through the following technical solutions:
[0009] A method for predicting the electrical energy consumption of an air source heat pump includes:
[0010] S1. Data Acquisition and Preprocessing: The operating parameters and meteorological environmental parameters of the air source heat pump are acquired through the data acquisition unit and preprocessed to form a time series dataset with a unified time reference.
[0011] S2. Frost Zone Identification and Quantification: Based on meteorological environmental parameters in the time series dataset, a frost susceptibility index is constructed. Based on the frost susceptibility index, sensitive operating zones for frost and defrosting conditions are identified. Combined with the virtual frost accumulation, an enhanced frost risk quantification sequence is generated.
[0012] S3. Defrosting detection and energy consumption decomposition: Identify and mark the start and end times of defrosting events based on energy consumption gradient values; using the steady-state power before the defrosting event as the baseline, decompose the total power consumption in each prediction period into a steady-state heating base energy consumption component and a defrosting pulse additional energy consumption component.
[0013] S4. Dual-channel prediction training: Construct a dual-channel prediction structure and train it to deduce the steady-state heating basic energy consumption component and the defrost pulse additional energy consumption component respectively.
[0014] S5. Prediction Fusion and Confidence Output: The prediction results of the two channels are superimposed and fused to obtain the total power consumption prediction value and calculate the prediction confidence. A parameter online update mechanism is established in combination with the rolling prediction accuracy index.
[0015] Furthermore, the specific operation steps of S2 are as follows:
[0016] Outdoor dry-bulb temperature, outdoor relative humidity, and outdoor wind speed were extracted from the time-series dataset. Frost susceptibility index was constructed using Gaussian temperature driving factor, humidity supply factor, and wind speed correction factor.
[0017] Based on the comparison between the frost susceptibility index and the preset threshold, sensitive operating ranges and normal operating ranges are marked respectively;
[0018] The frost susceptibility index within the sensitive operating range is smoothed using a sliding window mean to generate a quantified frost risk sequence.
[0019] Simultaneously, starting from the defrosting end time, the virtual frost accumulation is obtained by combining the frost susceptibility index and the operation offset correction term constructed based on the compressor exhaust temperature deviation.
[0020] If the virtual frost layer accumulation exceeds the preset threshold, it is determined to be a high risk of defrosting, and the virtual frost layer accumulation is reset after defrosting is completed.
[0021] The virtual frost accumulation is normalized and spliced with the frost risk quantification sequence at each time step to form an enhanced frost risk quantification sequence.
[0022] Furthermore, the specific operating steps for S3 are as follows:
[0023] The instantaneous power sequence of the compressor is extracted from the time series dataset, and the energy consumption gradient value is obtained based on the average instantaneous power of the compressor in adjacent windows by sliding along the time axis with a fixed window width.
[0024] Based on the energy consumption gradient value, suspected defrosting trigger points are marked. The pulse waveform of the suspected defrosting trigger points is verified. If a single peak pulse waveform with a rise followed by a fall is formed, it is determined that the energy consumption change is caused by the defrosting action. The start and end times of the corresponding defrosting event are recorded simultaneously. If no such waveform is formed, it is determined that the corresponding power change is caused by external interference factors. The corresponding suspected defrosting trigger point and related time period are excluded from the defrosting event labeling set.
[0025] Using the steady-state operating power before defrosting as a benchmark, the arithmetic mean of the instantaneous power of the compressor during the running period before the start of defrosting is selected as the steady-state baseline power. The additional energy consumption exceeding the steady-state operation during the defrosting process is calculated moment by moment, and the additional energy consumption of a single defrosting pulse is obtained by summing them up.
[0026] The additional energy consumption of a single defrost pulse for all defrost events within the predicted period is statistically analyzed to obtain the overall additional energy consumption component of the defrost pulse. The difference between this component and the total electricity consumption is then calculated to obtain the steady-state heating base energy consumption component.
[0027] Furthermore, the specific operation steps of S4 include:
[0028] The first prediction channel takes the steady-state heating basic energy consumption component as the prediction target and uses outdoor dry-bulb temperature, outdoor relative humidity, indoor set temperature, building envelope heat load index, and heat pump continuous operation duration as input feature vectors. It completes training based on the time-series regression mapping relationship oriented towards steady-state heating basic energy consumption. The training samples only select sample data from the normal operating range and steady-state heating operation sample data after removing the interference of defrost events. The training samples are steady-state heating operation sample data after removing the abnormal changes in energy consumption pulses caused by defrost events.
[0029] The second prediction channel uses the additional energy consumption component of the defrost pulse as the final prediction target. It adopts a two-level cascaded structure of defrost probability inference unit and defrost energy consumption amplitude inference unit for prediction. The first and second prediction channels are trained independently based on their respective training samples.
[0030] Furthermore, the specific operation steps of S4 also include:
[0031] The defrost probability simulation unit uses the frost susceptibility index and the normalized virtual frost accumulation as core inputs, and calculates the defrost trigger probability by combining the outdoor dry bulb temperature and outdoor relative humidity.
[0032] The defrost energy consumption amplitude extrapolation unit uses the historical single defrost pulse additional energy consumption as the fitting label, and the virtual frost layer accumulation before each defrost event and the continuous heating operation duration before defrost as the basic input features. At the same time, auxiliary input features are introduced to construct a time-series regression mapping relationship for defrost energy consumption amplitude and output the single defrost pulse additional energy consumption amplitude.
[0033] The expected number of defrost cycles is obtained by combining the historical average defrost frequency with the actual duration of the target period within the sensitive operating range.
[0034] The expected value of additional energy consumption of defrost pulse is obtained based on the amplitude of additional energy consumption of a single defrost pulse, the expected number of defrost cycles, and the defrost trigger probability.
[0035] During training, a near-critical oversampling strategy is used to generate synthetic positive samples, which are then added to the set of positive samples for defrosting events. The positive and negative samples are then adjusted to a preset balance ratio before training begins.
[0036] After the first and second prediction channels have been trained, cross-backtracking verification is performed.
[0037] Furthermore, the specific steps of the cross-checking are as follows:
[0038] The defrosting period data extrapolated from the second prediction channel in the historical sensitive operating range is input back into the first prediction channel to obtain the corresponding steady-state heating baseline energy consumption value and compare it with the actual steady-state energy consumption benchmark value to obtain the deviation range.
[0039] If the deviation exceeds a preset threshold, a linear correction coefficient is constructed based on the deviation to apply a downward correction to the output of the first prediction channel.
[0040] In the real-time prediction phase, the defrost trigger probability is updated based on the number of sampling cycles accumulated after the defrost is completed and the amount of virtual frost layer accumulated.
[0041] When the virtual frost layer accumulation monotonically increases for multiple consecutive sampling periods and the rate exceeds the historical average rate, the defrosting trigger probability is adjusted upwards.
[0042] The corrected defrost trigger probability is replaced with the original defrost trigger probability and substituted into the calculation process of the expected value of defrost additional energy consumption to complete the final deduction of the defrost pulse additional energy consumption for the target period.
[0043] Furthermore, the specific operation steps of S5 are as follows:
[0044] The total electricity consumption prediction value is obtained by time-series superposition and fusion of the energy consumption prediction values of the two channels;
[0045] The forecast period is determined by interval, and the corresponding energy consumption results are selected and summarized based on the normal and sensitive operating intervals respectively.
[0046] The prediction confidence level is obtained based on the frost susceptibility index and the fluctuation of virtual frost accumulation.
[0047] Based on the comparison between the prediction confidence level and the lower confidence level threshold, the prediction result is marked and an early warning is issued, or the predicted power consumption value and the corresponding confidence level are directly output.
[0048] A rolling comparison is made between the predicted energy consumption values and the measured energy consumption values for the completed forecast period, and the rolling forecast accuracy index is calculated.
[0049] If the rolling prediction accuracy index reaches the degradation trigger condition, the online parameter update mechanism is triggered. The measured data and labeled results are retrieved to perform incremental training on both channels. The gradient descent method is used to iteratively update the parameters until the rolling prediction accuracy index falls back to the accuracy degradation threshold. During the update process, historical parameter backups are retained, and each parameter update information is recorded to form a parameter update log.
[0050] Compared with the prior art, the beneficial effects of the present invention are:
[0051] This invention collects and preprocesses operating and meteorological parameters of air source heat pumps to calculate frost susceptibility indicators and identify frost intervals using virtual frost accumulation. It then verifies and detects defrosting events using energy consumption gradients and pulse waveforms, decomposing steady-state heating baseline energy consumption and defrosting pulse additional energy consumption. A dual-channel prediction training model is constructed to deduce the two types of energy consumption separately. Near-critical oversampling is used to address sample imbalance, and cross-checking is employed to optimize the training effect. The prediction results are fused and confidence levels are calculated. An online parameter update mechanism is established, combined with rolling prediction accuracy indicators, to achieve dynamic optimization of prediction parameters and iterative prediction channel iteration. This significantly improves the accuracy and reliability of air source heat pump electricity consumption prediction, effectively quantifies the additional energy consumption caused by frost and defrosting, avoids measurement biases in traditional methods, and enhances the engineering practicality and adaptability of energy consumption prediction results. Attached Figure Description
[0052] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0054] Example:
[0055] like Figure 1As shown, a method for predicting the power consumption of an air source heat pump includes data acquisition and preprocessing, frosting interval identification and quantification, defrosting detection and energy consumption decomposition, dual-channel prediction training, and prediction fusion and confidence output.
[0056] S1. Data Acquisition and Preprocessing: The data acquisition unit deployed on the air source heat pump system acquires the operating parameters and meteorological environmental parameters of the air source heat pump. The operating parameters include the instantaneous power of the compressor, the inlet and outlet temperatures of the outdoor heat exchanger, the compressor exhaust temperature, the suction pressure, and the exhaust pressure. The meteorological environmental parameters include the outdoor dry-bulb temperature, the outdoor relative humidity, and the outdoor wind speed. The acquired data is continuously collected at a fixed sampling period and preprocessed by outlier removal, linear interpolation of missing values, and timestamp alignment to form a time-series dataset with a unified time reference.
[0057] S2. Frost Zone Identification and Quantization:
[0058] Outdoor dry-bulb temperature, outdoor relative humidity, and outdoor wind speed are extracted from the time-series dataset time-by-time. At each sampling time, a Gaussian temperature driving factor is constructed based on the degree of deviation of the outdoor dry-bulb temperature from the center value of the high-frost-occurrence temperature. The positive deviation of the outdoor relative humidity exceeding the preset critical humidity threshold for frosting is used as the humidity supply factor, and the weakening effect of outdoor wind speed on the boundary layer of the outdoor heat exchanger fins is used as the wind speed correction factor. These parameters are then calculated using the formula... Calculations were performed to obtain The susceptibility index to frost formation at any given time, among which, Indicates the current moment. This represents the Gaussian temperature driving factor. This represents the natural exponential function. This represents the Gaussian decay coefficient at temperature. express outdoor dry-bulb temperature at any given time This indicates the central value of the temperature at which frost occurs most frequently. Indicates the humidity supply factor. express outdoor relative humidity at any given time This indicates the preset critical humidity threshold for frosting. This represents the wind speed correction factor. This represents the wind speed attenuation coefficient. express Outdoor wind speed at any given time;
[0059] The continuous period of time in which the frost susceptibility index value is greater than the preset threshold is marked as the sensitive operating range;
[0060] The time period when the frost susceptibility index value is less than the preset threshold is marked as the normal operating range;
[0061] Within the sensitive operating range, the frost susceptibility index values at multiple consecutive sampling times are averaged using a sliding window to smooth the data, eliminating high-frequency noise introduced by instantaneous fluctuations in meteorological parameters, and generating a frost risk quantification sequence to characterize the relative risk level of defrosting events at each time point.
[0062] Simultaneously, a virtual frost accumulation is constructed to characterize the frost growth and accumulation state on the surface of outdoor heat exchanger fins that cannot be directly measured.
[0063] At the end of the most recent defrosting event As the starting point of frost accumulation, from the beginning of accumulation to the current moment. At each discrete sampling time, the frost susceptibility index value is used as a proxy variable for the instantaneous frost rate, and the deviation of the compressor exhaust temperature from the compressor reference exhaust temperature is introduced to construct the operation offset correction term.
[0064] Through formula Calculations were performed to obtain The amount of virtual frost accumulated at any given time, where, This represents the discrete time points traversed during the cumulative calculation. Indicates the first The corresponding frost susceptibility index at any given time. Indicates the data sampling time interval. This indicates the preset exhaust temperature correction factor. Indicates the first The actual discharge temperature of the compressor at any given time. This indicates the compressor's rated reference discharge temperature under design operating conditions. This represents the operating offset correction term constructed based on the deviation of exhaust temperature;
[0065] When the compressor discharge temperature is higher than the reference discharge temperature, the operating offset correction term decreases, which reduces the equivalent frosting rate accordingly; when the compressor discharge temperature is lower than the reference discharge temperature, the operating offset correction term increases, which increases the equivalent frosting rate accordingly.
[0066] When the virtual frost layer accumulation exceeds the preset frost layer accumulation threshold, it is determined that the current moment is in a high-risk state where defrosting is about to be triggered; after each confirmed defrosting event ends, the virtual frost layer accumulation is reset to zero and accumulation starts again;
[0067] The virtual frost accumulation is normalized based on the frost accumulation threshold to obtain the normalized virtual frost accumulation, which is then spliced with the frost risk quantification sequence time by time to form an enhanced frost risk quantification sequence.
[0068] S3. Defrosting Detection and Energy Consumption Breakdown:
[0069] Extract the compressor instantaneous power sequence from the preprocessed time-series dataset; set a fixed window width. The sliding window is moved point by point along the time axis with a step size of a single sampling period. At each window position, the arithmetic mean of the instantaneous power of the compressor in the current window and the previous adjacent window is calculated. The energy consumption gradient value is constructed based on the relative change of the instantaneous power average between two adjacent windows. The calculation formula is as follows: ,in, express The energy gradient value at time step [time]. This represents the arithmetic mean of the compressor's instantaneous power within the current time window. This represents the arithmetic mean of the compressor's instantaneous power within the previous time window;
[0070] When the energy consumption gradient value at any moment is less than or equal to the preset gradient change threshold, it is determined that the current compressor power is in a stable change state and there is no power change caused by defrosting action.
[0071] When the energy consumption gradient value at any moment exceeds the preset gradient change threshold, it is determined that there is an abnormal power jump at the current moment, and the current moment is marked as a suspected defrost trigger point.
[0072] After marking the suspected defrosting trigger point, proceed to the pulse waveform verification stage:
[0073] Within the preset observation window range corresponding to the suspected defrost trigger point, feature matching judgment is performed on the instantaneous power timing waveform of the compressor;
[0074] If the compressor's instantaneous power continuously rises along the time axis and reaches a local power maximum within a preset time, then continues to decline and approaches the power level before defrosting is triggered, and then stably falls back to the steady-state operating range, forming a complete pulse waveform with a single peak value that rises first and then falls, it is determined to be an energy consumption change caused by defrosting action, and the start and end times of the corresponding defrosting event are recorded simultaneously.
[0075] If a complete pulse waveform with a single peak and a rise followed by a fall is not formed, it is determined that the corresponding power change is caused by equipment operation disturbance or external interference factors, and the corresponding suspected defrosting trigger point and related time period are excluded from the defrosting event label set.
[0076] After completing the defrost event identification, the energy consumption components are decomposed based on the steady-state operating power before defrost.
[0077] During the confirmed defrosting event period Within this range, the arithmetic mean of the instantaneous power of the compressor during a stable operating period prior to the start of defrosting is selected as the steady-state baseline power. Steady-state baseline power represents the power level of the heat pump during normal heating operation when there is no frost and no defrosting action is performed;
[0078] Based on the steady-state baseline power, the additional energy consumption exceeding steady-state operation during defrosting is calculated time-by-time, and the cumulative additional energy consumption of a single defrosting pulse is obtained. The calculation formula is as follows: ,in, This indicates the additional energy consumption of a single defrost pulse. Indicates the first The instantaneous power of the compressor at that moment;
[0079] For each complete time period to be predicted, all confirmed defrosting events within the time period are statistically analyzed, and the additional energy consumption of each defrosting pulse corresponding to each defrosting event is accumulated to obtain the overall additional energy consumption component of the defrosting pulse within the complete time period to be predicted.
[0080] Using the total electrical energy consumption of the heat pump during the complete period to be predicted as a benchmark, and calculating the difference between it and the additional energy consumption component of the overall defrost pulse during the corresponding period, the basic energy consumption component generated by the air source heat pump only performing steady-state heating is obtained.
[0081] S4, Dual-channel prediction training:
[0082] A first prediction channel is constructed for the steady-state heating basic energy consumption component, and regression prediction is performed using normal operating characteristics as input.
[0083] A second prediction channel is constructed for the additional energy consumption component of the defrost pulse. The enhanced frost risk quantification sequence and the virtual frost accumulation are used as the core input features to complete the deduction of the defrost trigger probability and the deduction of the additional energy consumption amplitude of a single defrost pulse, respectively. The two channels are trained independently based on their respective training samples.
[0084] The construction process of the first prediction channel is as follows:
[0085] The first prediction channel takes the basic energy consumption component generated by steady-state heating as the prediction target, and uses outdoor dry-bulb temperature, outdoor relative humidity, indoor set temperature, building envelope heat load index, and heat pump continuous operation time as input feature vectors. It completes training based on the time-series regression mapping relationship for steady-state heating basic energy consumption. The time-series regression mapping relationship refers to the mapping form pre-constructed based on time-series data to describe the stable correspondence between the input feature vector and the target predicted value.
[0086] During the selection of training samples, only sample data from the normal operating range and steady-state heating operation sample data after removing the interference of defrost events are selected, so that the training samples of the first prediction channel do not contain energy consumption pulse changes caused by defrost events, and ensure that the established time series regression mapping relationship only reflects the energy consumption change law under steady-state heating conditions.
[0087] The construction process of the second prediction channel is as follows:
[0088] The second prediction channel takes the additional energy consumption component of the defrost pulse as the final prediction target and adopts a two-stage series structure of defrost probability deduction unit and defrost energy consumption amplitude deduction unit.
[0089] The defrost probability extrapolation unit uses the frost susceptibility index and the normalized virtual frost accumulation in the enhanced frost risk quantification sequence as core inputs, combined with outdoor dry-bulb temperature and outdoor relative humidity, and calculates the result using the formula... Obtain the defrost trigger probability at the target sampling time. ,in, This represents the Sigmoid activation function. This represents the normalized virtual frost accumulation. This represents the bias term, which is trained synchronously with the weight parameters using gradient descent and is used to adjust the baseline value of the linear combination. These represent the influence weighting factors of frost susceptibility index, normalized virtual frost accumulation, outdoor dry-bulb temperature, and outdoor relative humidity, respectively.
[0090] The defrost energy consumption amplitude extrapolation unit uses the additional energy consumption of a single defrost pulse corresponding to each historically labeled defrost event as the fitting label, and the virtual frost accumulation before each defrost event and the continuous heating operation duration before defrost as the basic input features. At the same time, outdoor ambient temperature and outdoor relative humidity are introduced as auxiliary input features to construct a time-series regression mapping relationship for defrost energy consumption amplitude. Based on the time-series regression mapping relationship for defrost energy consumption amplitude, the unit outputs the additional energy consumption amplitude of a single defrost pulse.
[0091] Extract defrosting event records within the sensitive operating range from historical operating data, count the number of defrosting events occurring within a unit of time, and obtain the historical average defrosting frequency; based on the actual duration of the target time period within the sensitive operating range, multiply the historical average defrosting frequency by the duration of the sensitive operating range in the target time period to obtain the expected number of defrostings in the target time period.
[0092] Through formula The expected value of additional energy consumption for defrosting pulses within the target time period is obtained. ,in, The projected value representing the additional energy consumption amplitude of a single defrost pulse is derived by the defrost energy consumption amplitude projection unit through regression mapping, combining the virtual frost accumulation before defrost and the continuous heating operation duration before defrost. Indicates the expected number of defrost cycles within the target time period;
[0093] During training, the number of samples corresponding to defrost events in the historical runtime data is less than the number of samples in the regular heating runtime, resulting in an unbalanced distribution of sample categories. Therefore, a near-critical oversampling strategy based on the accumulation of virtual frost layers is adopted.
[0094] Samples with virtual frost accumulation between 80% and 100% of the frost accumulation threshold but which did not actually trigger defrosting were selected as near-critical samples. Using the near-critical samples as seeds, a small positive perturbation with uniform distribution was applied to the single feature dimension of virtual frost accumulation. The perturbation amplitude did not exceed 5% of the frost accumulation threshold. Synthetic positive samples were generated and added to the set of positive samples of defrosting events. The ratio of positive to negative samples was adjusted to a preset balance ratio before training was carried out.
[0095] After the first and second prediction channels have been trained, perform cross-backtracking verification:
[0096] The defrosting period data derived from the second prediction channel on historical sensitive operating intervals is input back into the first prediction channel to obtain the steady-state heating baseline energy consumption value output by the first prediction channel for the corresponding period.
[0097] The output steady-state heating baseline energy consumption value is compared with the actual steady-state energy consumption baseline value after defrosting interference to obtain the deviation range;
[0098] When the deviation exceeds a preset threshold, a linear correction coefficient is constructed based on the deviation, and a downward correction is applied to the output of the first prediction channel in the corresponding time period;
[0099] When the deviation is lower than the preset threshold, the original output value of the first prediction channel in the corresponding time period is retained, no correction operation is performed, and the original prediction result remains unchanged.
[0100] In the real-time prediction phase, the defrost trigger probability is updated based on the number of sampling cycles accumulated since the end of the most recent defrost event and the current virtual frost layer accumulation.
[0101] When the virtual frost accumulation remains monotonically increasing over multiple consecutive sampling periods, and the rate of increase exceeds the average rate of increase before historical defrost events are triggered, the formula is used to determine the outcome. The defrost trigger probability has been adjusted upwards. express The corrected defrost trigger probability at any given time. This represents the minimum value constraint function. This indicates that the preset probability is adjusted upwards by a correction factor. The variable indicates that the rate of increase exceeds the limit. When the rate of increase of the virtual frost layer accumulation exceeds the average rate of increase before the historical defrost event is triggered, the value is 1. When the rate of increase of the virtual frost layer accumulation does not exceed the average rate of increase before the historical defrost event is triggered, the value is 0.
[0102] The original defrost trigger probability is replaced with the corrected defrost trigger probability and substituted into the calculation process of the expected value of defrost additional energy consumption to complete the final deduction of the defrost pulse additional energy consumption for the target period.
[0103] S5. Predictive fusion and confidence output:
[0104] The steady-state heating basic energy consumption prediction value output by the first prediction channel is time-series superimposed and fused with the defrost pulse additional energy consumption expectation value output by the second prediction channel to obtain the total power consumption prediction value of the air source heat pump within the complete prediction period.
[0105] For the period to be predicted, the interval segmentation judgment is performed. If the time is determined to be a normal operating interval, the steady-state heating basic energy consumption prediction value is directly used as the energy consumption result of the corresponding segment. If the time is determined to be a sensitive operating interval, the time series superposition and fusion result is used as the energy consumption result of the corresponding segment. After the calculation is completed for each time period, the total power consumption prediction value for the whole time period is obtained by summing them up.
[0106] Based on the fluctuation of the current frost susceptibility index and the virtual frost accumulation, the formula is used to... Calculate the prediction confidence level ,in, This represents a smoothing constant set to prevent the denominator from being zero. Variance calculation function Indicates the number of statistical sampling periods. This indicates the preset frost accumulation threshold. This indicates the relative fluctuation in the amount of virtual frost accumulation. This indicates the confidence decay weight corresponding to the susceptibility index to frost formation. The confidence decay weight represents the relative fluctuation of the accumulated virtual frost layer.
[0107] When the prediction confidence level is lower than the preset confidence level threshold, the current prediction result is marked as a low confidence prediction at the output end and an early warning is triggered. At the same time, the corresponding upper and lower fluctuation ranges are attached to the output of the final electricity consumption prediction value to intuitively reflect the uncertainty range of the prediction result.
[0108] When the prediction confidence level is higher than the preset confidence level threshold, the current energy consumption prediction result is determined to be of high confidence level. There is no need to trigger an early warning signal. The final electricity consumption prediction value and the corresponding confidence level are directly output, and the prediction parameters and confidence level data for the corresponding time period are recorded simultaneously.
[0109] For multiple consecutive completed forecast periods, the predicted final electricity consumption value for each period is compared with the corresponding measured energy consumption value on a rolling basis using a formula. The rolling forecast accuracy index is obtained through calculation. ,in, This represents the total number of forecast periods that have been completed. Indicates the first Total electricity consumption forecast for each time period Indicates the first Measured total electricity consumption for each time period;
[0110] When the rolling prediction accuracy index remains above the preset accuracy degradation threshold for multiple consecutive sampling periods, or when the magnitude of a single exceedance of the accuracy degradation threshold reaches the preset deviation upper limit, the prediction accuracy is determined to meet the degradation trigger condition, triggering the online parameter update mechanism.
[0111] Retrieve complete measured operating data and corresponding annotation results within a preset time period. The measured operating data includes the operating parameters of the air source heat pump and meteorological environmental parameters. The annotation results include the defrost event annotations for the corresponding time period, the additional energy consumption annotations for a single defrost pulse, the basic energy consumption annotations for steady-state heating, and the measured annotations for total electricity consumption.
[0112] Incremental training is performed on the first and second prediction channels respectively:
[0113] For the first prediction channel, the samples of the normal operating range and the steady-state operating samples after removing defrost interference in the sensitive operating range from the retrieved measured data are used as incremental training samples. The measured value of steady-state heating basic energy consumption is used as the fitting target. The original time-series regression mapping relationship of the first prediction channel is adjusted, the weight parameters corresponding to the input features are updated, and the mapping accuracy between features and steady-state energy consumption is optimized.
[0114] For the second prediction channel, defrost event samples and near-critical samples from the retrieved measured data are used as incremental training samples. Incremental training is performed on the defrost probability inference unit and the defrost energy consumption amplitude inference unit respectively: For the defrost probability inference unit, the weighting factors and bias terms corresponding to the frost susceptibility index, normalized virtual frost accumulation, outdoor dry-bulb temperature, and outdoor relative humidity are updated to optimize the inference accuracy of defrost trigger probability; For the defrost energy consumption amplitude inference unit, the time-series regression mapping relationship between virtual frost accumulation, continuous heating operation time before defrost, environmental parameters and the additional energy consumption amplitude of a single defrost pulse is updated to correct the amplitude inference bias; At the same time, if the addition of new defrost event samples causes an imbalance in the ratio of positive and negative samples, a near-critical oversampling strategy based on virtual frost accumulation is used to supplement and synthesize positive samples to maintain sample balance before completing incremental training.
[0115] During incremental training, the gradient descent method is used to iteratively update the internal weights and mapping parameters of the channels round by round. After each round of parameter update, the rolling prediction accuracy index of the total power consumption prediction value of the dual-channel output after the update and the actual value of the corresponding time period is immediately calculated and compared with the preset accuracy degradation threshold. If the rolling prediction accuracy index still exceeds the accuracy degradation threshold, incremental training and parameter update are continued until the rolling prediction accuracy index falls back to within the accuracy degradation threshold. Then, parameter update is stopped and the latest channel parameters are saved as the benchmark parameters for subsequent predictions.
[0116] During the online parameter update process, historical parameter backups are retained. If abnormal fluctuations in prediction accuracy occur after the update, a parameter rollback mechanism can be triggered to restore the most recent stable parameter state, thus avoiding prediction failure caused by improper parameter updates. At the same time, the time, update magnitude, and corresponding accuracy changes of each parameter update are recorded to form a parameter update log, providing data support for subsequent threshold optimization and parameter adjustment.
[0117] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting the electrical energy consumption of an air source heat pump, characterized in that, include: S1. Data Acquisition and Preprocessing: The operating parameters and meteorological environmental parameters of the air source heat pump are acquired through the data acquisition unit and preprocessed to form a time series dataset with a unified time reference. S2. Frost Zone Identification and Quantification: Based on meteorological environmental parameters in the time series dataset, a frost susceptibility index is constructed. Based on the frost susceptibility index, sensitive operating zones for frost and defrosting conditions are identified. Combined with the virtual frost accumulation, an enhanced frost risk quantification sequence is generated. S3. Defrosting detection and energy consumption decomposition: Identify and mark the start and end times of defrosting events based on energy consumption gradient values; using the steady-state power before the defrosting event as the baseline, decompose the total power consumption in each prediction period into a steady-state heating base energy consumption component and a defrosting pulse additional energy consumption component. S4. Dual-channel prediction training: Construct a dual-channel prediction structure and train it to deduce the steady-state heating basic energy consumption component and the defrost pulse additional energy consumption component respectively. S5. Prediction Fusion and Confidence Output: The prediction results of the two channels are superimposed and fused to obtain the total power consumption prediction value and calculate the prediction confidence. A parameter online update mechanism is established in combination with the rolling prediction accuracy index.
2. The method for predicting the electrical energy consumption of an air source heat pump according to claim 1, characterized in that, The specific operation steps of S2 are as follows: Outdoor dry-bulb temperature, outdoor relative humidity, and outdoor wind speed were extracted from the time-series dataset. Frost susceptibility index was constructed using Gaussian temperature driving factor, humidity supply factor, and wind speed correction factor. Based on the comparison between the frost susceptibility index and the preset threshold, sensitive operating ranges and normal operating ranges are marked respectively; The frost susceptibility index within the sensitive operating range is smoothed using a sliding window mean to generate a quantified frost risk sequence. Simultaneously, starting from the defrosting end time, the virtual frost accumulation is obtained by combining the frost susceptibility index and the operation offset correction term constructed based on the compressor exhaust temperature deviation. If the virtual frost layer accumulation exceeds the preset threshold, it is determined to be a high risk of defrosting, and the virtual frost layer accumulation is reset after defrosting is completed. The virtual frost accumulation is normalized and spliced with the frost risk quantification sequence at each time step to form an enhanced frost risk quantification sequence.
3. The method for predicting the electrical energy consumption of an air source heat pump according to claim 1, characterized in that, The specific operating steps for S3 are as follows: The instantaneous power sequence of the compressor is extracted from the time series dataset, and the energy consumption gradient value is obtained based on the average instantaneous power of the compressor in adjacent windows by sliding along the time axis with a fixed window width. Based on the energy consumption gradient value, suspected defrosting trigger points are marked. The pulse waveform of the suspected defrosting trigger points is verified. If a single peak pulse waveform with a rise followed by a fall is formed, it is determined that the energy consumption change is caused by the defrosting action. The start and end times of the corresponding defrosting event are recorded simultaneously. If no such waveform is formed, it is determined that the corresponding power change is caused by external interference factors. The corresponding suspected defrosting trigger point and related time period are excluded from the defrosting event labeling set. Using the steady-state operating power before defrosting as a benchmark, the arithmetic mean of the instantaneous power of the compressor during the running period before the start of defrosting is selected as the steady-state baseline power. The additional energy consumption exceeding the steady-state operation during the defrosting process is calculated moment by moment, and the additional energy consumption of a single defrosting pulse is obtained by summing them up. The additional energy consumption of a single defrost pulse for all defrost events within the predicted period is statistically analyzed to obtain the overall additional energy consumption component of the defrost pulse. The difference between this component and the total electricity consumption is then calculated to obtain the steady-state heating base energy consumption component.
4. The method for predicting the electrical energy consumption of an air source heat pump according to claim 1, characterized in that, The specific operation steps of S4 include: The first prediction channel takes the steady-state heating basic energy consumption component as the prediction target and uses outdoor dry-bulb temperature, outdoor relative humidity, indoor set temperature, building envelope heat load index, and heat pump continuous operation duration as input feature vectors. It completes training based on the time-series regression mapping relationship oriented towards steady-state heating basic energy consumption. The training samples only select sample data from the normal operating range and steady-state heating operation sample data after removing the interference of defrost events. The training samples are steady-state heating operation sample data after removing the abnormal changes in energy consumption pulses caused by defrost events. The second prediction channel uses the additional energy consumption component of the defrost pulse as the final prediction target. It adopts a two-level cascaded structure of defrost probability inference unit and defrost energy consumption amplitude inference unit for prediction. The first and second prediction channels are trained independently based on their respective training samples.
5. The method for predicting the power consumption of an air source heat pump according to claim 4, characterized in that, The specific operation steps of S4 also include: The defrost probability simulation unit uses the frost susceptibility index and the normalized virtual frost accumulation as core inputs, and calculates the defrost trigger probability by combining the outdoor dry bulb temperature and outdoor relative humidity. The defrost energy consumption amplitude extrapolation unit uses the historical single defrost pulse additional energy consumption as the fitting label, and the virtual frost layer accumulation before each defrost event and the continuous heating operation duration before defrost as the basic input features. At the same time, auxiliary input features are introduced to construct a time-series regression mapping relationship for defrost energy consumption amplitude and output the single defrost pulse additional energy consumption amplitude. The expected number of defrost cycles is obtained by combining the historical average defrost frequency with the actual duration of the target period within the sensitive operating range. The expected value of additional energy consumption of defrost pulse is obtained based on the amplitude of additional energy consumption of a single defrost pulse, the expected number of defrost cycles, and the defrost trigger probability. During training, a near-critical oversampling strategy is used to generate synthetic positive samples, which are then added to the set of positive samples for defrosting events. The positive and negative samples are then adjusted to a preset balance ratio before training begins. After the first and second prediction channels have been trained, cross-backtracking verification is performed.
6. The method for predicting the electrical energy consumption of an air source heat pump according to claim 5, characterized in that, The specific steps for the cross-backtracking verification are as follows: The defrosting period data extrapolated from the second prediction channel in the historical sensitive operating range is input back into the first prediction channel to obtain the corresponding steady-state heating baseline energy consumption value and compare it with the actual steady-state energy consumption benchmark value to obtain the deviation range. If the deviation exceeds a preset threshold, a linear correction coefficient is constructed based on the deviation to apply a downward correction to the output of the first prediction channel. In the real-time prediction phase, the defrost trigger probability is updated based on the number of sampling cycles accumulated after defrost ends and the accumulated amount of virtual frost layer. When the virtual frost layer accumulation monotonically increases for multiple consecutive sampling periods and the rate exceeds the historical average rate, the defrosting trigger probability is adjusted upwards. The corrected defrost trigger probability is replaced with the original defrost trigger probability and substituted into the calculation process of the expected value of defrost additional energy consumption to complete the final deduction of the defrost pulse additional energy consumption for the target period.
7. The method for predicting the electrical energy consumption of an air source heat pump according to claim 1, characterized in that, The specific operation steps of S5 are as follows: The total electricity consumption prediction value is obtained by time-series superposition and fusion of the energy consumption prediction values of the two channels; The forecast period is determined by interval, and the corresponding energy consumption results are selected and summarized based on the normal and sensitive operating intervals respectively. The prediction confidence level is obtained based on the frost susceptibility index and the fluctuation of virtual frost accumulation. Based on the comparison between the prediction confidence level and the lower confidence level threshold, the prediction result is marked and an early warning is issued, or the predicted power consumption value and the corresponding confidence level are directly output. A rolling comparison is made between the predicted energy consumption values and the measured energy consumption values for the completed forecast period, and the rolling forecast accuracy index is calculated. If the rolling prediction accuracy index reaches the degradation trigger condition, the online parameter update mechanism is triggered. The measured data and labeled results are retrieved to perform incremental training on both channels. The gradient descent method is used to iteratively update the parameters until the rolling prediction accuracy index falls back to the accuracy degradation threshold. During the update process, historical parameter backups are retained, and each parameter update information is recorded to form a parameter update log.