Greenhouse heat pump unit load pre-adjustment method based on outdoor weather forecast

Abstract

The present application relates to a greenhouse heat pump unit load pre-adjustment method based on outdoor weather forecast, which compares the local micro-weather data collected by the low-cost micro-weather station array with the public weather forecast data, dynamically generates the hourly forecast error vector, and gives the forecast reliability label. A ternary coupled thermal inertia benchmark library is established combining the greenhouse structure, crops and climate parameters to realize the rapid retrieval of thermal response parameters. The main channel weather time series characteristics, auxiliary channel error information and thermal inertia parameters are fused, and the compensation weight is dynamically adjusted through the gating mechanism to implement the feedforward correction of the load prediction value under the physical feasible region. Finally, the heat pump start-stop control instruction with fault tolerance is generated, and the error model is dynamically optimized combined with the execution feedback. The scheme can effectively improve the accuracy of greenhouse heat load prediction and the stability of control, and enhance the adaptability and robustness of the system.

Description

Technical Field

[0001] This invention relates to the field of intelligent regulation and control and meteorological load prediction technology in facility agriculture, and in particular to a method for pre-adjusting the load of greenhouse heat pump units based on outdoor weather forecasts. Background Technology

[0002] Currently, the field of intelligent temperature control in facility agriculture widely adopts greenhouse heat pump load forecasting methods based on meteorological data and thermal models to automate the regulation of environmental parameters and energy supply for greenhouse groups. Mainstream technologies include: heat balance models driven by deterministic input of public meteorological forecast data; simplified thermal inertia models and big data-driven load time series regression models; and integrated deep learning and multi-factor mapping prediction algorithms that have emerged in recent years. These methods generally utilize macro-meteorological elements such as air temperature, humidity, and wind speed, combined with greenhouse building thermal parameters (such as envelope type, covering material characteristics, and heat storage medium arrangement), to periodically output short-term or ultra-short-term heat load forecasts, providing basic data support for the start-up and shutdown control, power grading, and energy-saving scheduling of greenhouse heat pump units.

[0003] In current application solutions, regional weather forecast data released daily by meteorological departments in various regions is often input into the management platform via API, FTP, or other means. High-precision field-measured data is usually difficult to deploy on a large scale due to limitations in equipment and cost, especially in small and medium-sized greenhouse clusters. To improve the robustness and safety of the system under extreme weather conditions, some solutions adopt extreme weather event classification switching and physical model preset extreme condition parameters, but most rely on statistical compensation methods or historical extreme case backtracking optimization.

[0004] With the development of deep learning methods, some existing load forecasting models attempt to incorporate neural network structures (such as LSTM, GRU, etc.) to fuse multi-source data, achieving the ability of the output curve to adaptively model historical loads, meteorological sequences, and crop growth stages. However, in actual deployment, these models typically have two prominent problems: First, they directly input data provided by meteorological forecasting services as exogenous "deterministic variables." Even with posterior residual compensation, they can only correct for "occurred / observed" biases after the fact, making it difficult to capture and eliminate feedforward errors caused by systematic drift or abrupt changes in forecast sources in a timely manner. Second, under extreme weather conditions, the models often rely heavily on historical extreme samples, limiting their training and generalization capabilities. Once new features emerge in extreme event types or the number of samples is insufficient, the models are prone to prediction distortion, over-adjustment, or even loss of control over the safety boundary.

[0005] Representative technologies such as "a method for predicting the heat load of solar greenhouses based on meteorological data," "a greenhouse cluster energy consumption prediction platform combining multi-factor regression," and "a dynamic response model for heat load based on LSTM" are applicable to scenarios involving the main energy management of facility agriculture and the optimization and scheduling of regional energy supply. These solutions have a certain degree of scalability, and after parameter optimization, they can meet the daily energy consumption planning needs under normal weather and trend-like fluctuations. However, in the complex scenario of "climate warning - extreme weather - dynamic information uncertainty," because it is impossible to measure in real time the linkage between the error characteristics of the meteorological source itself and the specific physical thermal inertia of the greenhouse, the model control output often exhibits "regulation lag, amplitude runaway, or cyclical oscillation," and may even misjudge transient events such as short-term low temperature / high humidity / sudden wind, leading to energy waste or regulation lag.

[0006] The aforementioned technologies generally suffer from the following prominent shortcomings: Meteorological inputs serve only as static deterministic data sources, failing to perceive, quantify, and dynamically compensate for the multidimensional distribution, abrupt change intensity, and directional characteristics of meteorological forecast errors in real time.

[0007] The thermal inertia parameters of building and crop systems are simplified or merged, and most models fail to characterize the spatiotemporal thermal response delay characteristics under different structures, materials, heat storage configurations and crop phenological stages.

[0008] The response to extreme weather events is highly dependent on historical samples or manual strategy switching, making it difficult to break away from the classification of extreme events and mapping of large sample sets, resulting in weak model generalization and cost reduction and efficiency improvement capabilities.

[0009] When model predictions are inaccurate, they can usually only be corrected through posterior statistical methods. It is impossible to achieve feedforward automatic suppression and correction based on the closed loop of real-time physical response state and forecast source error state, making it difficult to guarantee the physical interpretability and high safety of load forecasting and unit control.

[0010] When faced with marginalized small greenhouses or multi-greenhouse clusters, existing solutions struggle to achieve low computing power, high real-time performance, and environmental adaptation without relying on extreme training samples. They also exhibit weak adaptability to changes in forecast sources and climate anomalies during long-term operation.

[0011] Therefore, the industry urgently needs an innovative load forecasting method that can dynamically perceive the errors of weather forecast sources in real time and couple the error state with the physical and thermal inertia processes of the greenhouse building-crop system. This requires a physical fusion modeling of weather forecast errors and thermophysical response uncertainties through low-cost micro-weather sensor arrays, efficient dynamic error identification, real-time thermal inertia compensation, feedforward correction, and physical boundary constraints. This model should reflect the actual compensation effects of exogenous weather variations while ensuring that the model's input and output are always constrained by the inherent thermophysical nature of the greenhouse, significantly improving system robustness, interpretability, and long-term adaptive stability, especially in extreme weather changes and marginalized low-cost application scenarios. This is precisely the core technical deficiency and industry need that this invention aims to address. Summary of the Invention

[0012] This application provides a method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecasts, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0013] The load pre-adjustment method for greenhouse heat pump units based on outdoor weather forecasts provided in this application specifically includes: S1: Acquire real-time changes in local temperature and humidity, wind speed and public weather forecast data, and compare the real-time changes in local temperature and humidity and wind speed with the public weather forecast data to generate a forecast error vector sequence; S2: Perform confidence quantification based on the forecast error vector sequence to generate forecast confidence labels; S3: For greenhouse buildings with different structural types, different covering materials and different internal heat storage configurations, a set of thermal inertia parameters is constructed using the thermal response time constant, temperature hysteresis phase angle and load attenuation coefficient obtained from offline calibration; S4: Obtain conventional weather forecast elements, take the time series features of the conventional weather forecast elements as the main input, and take the forecast confidence label and the corresponding thermal inertia parameter set index as auxiliary input, perform feature fusion processing, and generate a fused feature vector; S5: Based on the abrupt change intensity of the prediction error vector in the fused feature vector, dynamically adjust the thermal inertia compensation weight to generate a control signal; S6: Use the control signal to perform advance deviation correction processing on the load change trend. When a sudden drop in temperature forecast is detected and the local measured temperature drop is slow, activate the thermal inertia delay response. Or when the forecast is continuously warmer and the local humidity rises sharply, call the crop transpiration correction coefficient to generate the heat load prediction value. S7: Based on the predicted heat load value, and combined with the error range constrained by the forecast confidence label and the thermal inertia parameter set, generate a heat pump unit preheating or precooling start / stop control command.

[0014] S8: Obtain the execution result of the preheating or precooling start-stop control command of the heat pump unit, and update the statistical distribution characteristics of the forecast error vector sequence based on the deviation between the execution result of the preheating or precooling start-stop control command of the heat pump unit and the subsequent actual meteorological data.

[0015] The load pre-adjustment method for greenhouse heat pump units based on outdoor weather forecasts provided in this application has the following beneficial effects: (1) To address the problems of poor robustness of traditional greenhouse heat pump load forecasting methods under extreme weather conditions and over-reliance on high-precision meteorological input or large-scale historical extreme sample training, this solution models meteorological forecast errors as observable state variables with spatiotemporal evolution characteristics and forms a closed-loop coupling mechanism in combination with the thermal dynamic characteristics of the greenhouse itself, which significantly improves the adaptability and stability of the forecasting model under meteorological change scenarios. Existing technologies usually treat meteorological data as deterministic input and only perform statistical corrections after forecasting, which is difficult to cope with the systematic deviations caused by short-term strong disturbances. However, this invention deploys a lightweight meteorological error sensing unit and uses a low-cost micro-meteorological station array to capture local microclimate changes in real time, generating a "forecast credibility label" with a time label, enabling the model to actively sense and quantify the local inaccuracy of the current forecast, thereby realizing a paradigm shift from "passively receiving forecasts" to "actively identifying errors", effectively overcoming the problem of load misjudgment caused by input distortion in traditional methods under extreme weather conditions such as cold waves, sudden rain, and strong winds.

[0016] (2) By constructing a greenhouse thermal inertia feature fingerprint database with structure-crop-climate ternary coupling and designing an error-inertia coupling gating module to achieve dual-channel feature fusion, this scheme realizes fine-grained response adjustment for different greenhouse types and operating states, which greatly improves the physical interpretability and adaptability of load forecasting. Unlike the technical path that relies on complex neural network structures or multi-model switching strategies, this invention establishes an indexable thermal inertia database based on offline calibrated physical parameters such as thermal response time constant, temperature lag phase angle and load attenuation coefficient. During online forecasting, this prior knowledge is introduced through an auxiliary channel, so that the model can quickly adapt to new greenhouses or seasonal change scenarios without extensive training. At the same time, the gating module adaptively adjusts the thermal inertia compensation weight according to the intensity of forecast error mutation—suppressing over-prediction in scenarios where the temperature drops sharply but the measured temperature rise is delayed, and enhancing the evaporation factor when the humidity increases suddenly and the latent heat load rises, thus realizing dynamic coordination correction of sensible heat and latent heat components, which significantly enhances the reliability and continuity of the forecast results under non-steady-state meteorological conditions.

[0017] (3) The proposed load prediction output form is a physical feasible domain boundary constrained by the error perception state and the thermal inertia response limit, rather than a confidence interval in the traditional statistical sense. This uncertainty boundary has a clear physical interpretation basis and can be directly used to formulate the preheating / precooling control strategy of the heat pump unit, thereby improving the response sensitivity and energy efficiency of the entire temperature control system. Since this boundary comprehensively considers the local meteorological error evolution trend and the greenhouse's own thermal buffering capacity, it avoids energy waste caused by conservative control or temperature fluctuations caused by aggressive adjustment. It is especially suitable for the small and medium-sized greenhouse clusters that are common in facility agriculture. Its low computing power requirement and parameter-free characteristics support long-term stable deployment on edge devices without relying on cloud computing or external risk indicator modeling, truly realizing the integrated architecture of "lightweight perception - mechanistic modeling - closed-loop control". Overall, this scheme constructs a new prediction paradigm that does not rely on extreme sample training, does not require complex network structures, has strong physical constraints and good generalizability, and provides reliable, efficient and economical technical support for intelligent control of greenhouse environment under extreme meteorological disturbances. Attached Figure Description

[0018] Figure 1 This is the main flowchart of the load pre-adjustment method for greenhouse heat pump units based on outdoor weather forecasts.

[0019] Figure 2 This is a sub-flowchart of a method for pre-adjusting the load of greenhouse heat pump units based on outdoor weather forecasts.

[0020] Figure 3 This is another sub-flowchart of the greenhouse heat pump unit load pre-adjustment method based on outdoor weather forecasts. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0023] like Figure 1As shown, this application provides a method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecasts, specifically including: S1: Acquire real-time changes in local temperature and humidity, wind speed and public weather forecast data, and compare the real-time changes in local temperature and humidity and wind speed with the public weather forecast data to generate a forecast error vector sequence; S2: Perform confidence quantification based on the forecast error vector sequence to generate forecast confidence labels; S3: For greenhouse buildings with different structural types, different covering materials and different internal heat storage configurations, a set of thermal inertia parameters is constructed using the thermal response time constant, temperature hysteresis phase angle and load attenuation coefficient obtained from offline calibration; S4: Obtain conventional weather forecast elements, take the time series features of the conventional weather forecast elements as the main input, and take the forecast confidence label and the corresponding thermal inertia parameter set index as auxiliary input, perform feature fusion processing, and generate a fused feature vector; S5: Based on the abrupt change intensity of the prediction error vector in the fused feature vector, dynamically adjust the thermal inertia compensation weight to generate a control signal; S6: Use the control signal to perform advance deviation correction processing on the load change trend. When a sudden drop in temperature forecast is detected and the local measured temperature drop is slow, activate the thermal inertia delay response. Or when the forecast is continuously warmer and the local humidity rises sharply, call the crop transpiration correction coefficient to generate the heat load prediction value. S7: Based on the predicted heat load value, and combined with the error range constrained by the forecast confidence label and the thermal inertia parameter set, generate a heat pump unit preheating or precooling start / stop control command.

[0024] S8: Obtain the execution result of the preheating or precooling start-stop control command of the heat pump unit, and update the statistical distribution characteristics of the forecast error vector sequence based on the deviation between the execution result of the preheating or precooling start-stop control command of the heat pump unit and the subsequent actual meteorological data.

[0025] Step S1: Acquire real-time changes in local temperature, humidity, and wind speed, as well as public weather forecast data. Compare the real-time changes in local temperature, humidity, and wind speed with the public weather forecast data to generate a forecast error vector sequence. Specifically, this includes: S1.1: Obtain the original local temperature and humidity signals and the original local wind speed micro-variation signals collected by a low-cost micro-weather station array deployed in the edge area of ​​the greenhouse cluster. Perform timestamp alignment and outlier removal processing on the original local temperature and humidity signals and the original local wind speed micro-variation signals to generate a standardized local micro-meteorological dataset.

[0026] The system acquires various sensor outputs from a low-cost micro weather station array deployed at the edge of the greenhouse cluster, including raw temperature signals and raw humidity signals to reflect the local thermal and humid environment, as well as raw wind speed micro-variation signals to characterize subtle changes in local airflow, forming a time-seriesd raw sensor data stream.

[0027] The original temperature signal, original humidity signal, and original wind speed micro-change signal are subjected to unified timestamp alignment processing. Based on the sampling frequency differences and clock drift of each sensor, the sampling point positions are allocated on a unified global time axis through an interpolation alignment algorithm to ensure that different types of signals form a strictly matched set of observation pairs in the time domain.

[0028] Outlier removal is performed on various types of signals after timestamp alignment. An outlier detection algorithm based on statistical tests is used to calculate the mean and standard deviation of each type of signal within a preset sliding window, and a confidence interval threshold is set. Data points falling outside the threshold are marked as outliers and removed.

[0029] A combination of physical model constraints and multi-channel consistency checks is used to verify the consistency of signals after outlier removal. Cross-validation is performed on the correlation between the abrupt changes in temperature and humidity signals and abnormal fluctuations in wind speed. If the physical correlation boundary is not met, the outliers are removed again.

[0030] The verified signals are standardized by numerical normalization to unify the dimensions and numerical ranges of different types of signals. Low-order bandpass filtering is used to suppress noise in the wind speed micro-variation signal. Finally, a standardized local micro-meteorological dataset that can be directly used as a basis for forecast error comparison is generated.

[0031] Through the above timestamp alignment and outlier removal processes, the raw sensor signals collected in the previous step are transformed into a standardized local micro-meteorological dataset that has temporal consistency, physical interpretability, and is suitable for downstream comparison calculations, thereby significantly improving the quality of meteorological signal data.

[0032] For example, in a greenhouse cluster scenario, a micro weather station array includes 20 nodes each for temperature, humidity, and wind speed sensors. The temperature sampling frequency is 1Hz, humidity is 0.5Hz, and wind speed is 2Hz. The maximum clock drift within the array is 0.8 seconds. For the collected temperature signals, linear interpolation alignment is performed over the interval t=0 to t=3600 seconds to match the temperature, humidity, and wind speed signals at a uniform time step of 1 second. The aligned temperature signals are then used to calculate the mean μ and standard deviation σ within a sliding window w=60 seconds. When the sampled values ​​satisfy |x Data points with μ|>3σ are considered abnormal and removed. During the verification phase, a physical correlation test is used to limit the temperature drop to no more than twice the humidity fluctuation, and the wind speed change rate |dv / dt| < 0.5 m / s². If these conditions are not met, the corresponding data points are removed. In the normalization phase, the temperature is mapped to the [0,1] interval, and a 0.1–1Hz bandpass filter is used to process the small wind speed variations to eliminate high-frequency noise. The final output standardized local micrometeorological dataset contains one triplet data point per second, representing the temperature, humidity, and filtered wind speed variations. This dataset can be directly used for public weather forecast comparison calculations in sub-step S1.2. In testing, this dataset significantly improved the stability of the downstream error vector.

[0033] S1.2: Based on the time window of the standardized local micro-meteorological dataset, the public meteorological forecast service interface is called to extract the public meteorological forecast service data within the corresponding spatiotemporal range. The public meteorological forecast service data includes forecast temperature data, forecast humidity data, and forecast wind speed data. The public meteorological forecast service data is subjected to interpolation and resampling processing to generate a synchronous public meteorological forecast sequence with the same time resolution as the standardized local micro-meteorological dataset.

[0034] Based on the time index of the local micrometeorological dataset after standardization in step S1.1, the original sequence data of forecast temperature, forecast humidity and forecast wind speed within the corresponding spatiotemporal range are obtained by calling the public weather forecast service interface that has been connected, ensuring that the interface retrieval parameters are strictly consistent with the spatial coordinates and time window of the local dataset.

[0035] Time series interpolation is performed on the forecast temperature data. Spline interpolation is used to generate continuous curves between the time nodes of the original forecast data. Samples are taken at the timestamp positions corresponding to the local dataset to make up for the difference in time resolution between the original forecast data and the local collected data.

[0036] The same interpolation process is applied to the predicted humidity data. Piecewise quadratic splines are used to maintain the smoothness of humidity changes. At the same time, out-of-range values ​​that do not conform to physical constraints are removed during the interpolation process to ensure the consistency between the humidity prediction and the actual physical environment.

[0037] When performing interpolation resampling on the forecast wind speed data, a wind speed mutation detection mechanism is introduced. When an abnormal gradient mutation is detected in the original wind speed sequence, a weighted moving average model is called to perform smoothing correction, so as to eliminate the impact of short-term extreme values ​​on the resampling accuracy.

[0038] The interpolated and resampled forecast temperature, humidity, and wind speed sequences are standardized to the same time resolution as the local micrometeorological dataset, and a synchronous public meteorological forecast sequence is constructed as the input for time-delay alignment processing.

[0039] By interpolating and resampling, the public weather forecast data corresponding to the standardized local micrometeorological dataset is transformed into a synchronous public weather forecast sequence with consistent time resolution and effective physical constraints, so as to achieve time base consistency for subsequent short-delay matching and deviation calculation.

[0040] For example, a low-cost micro-weather station array deployed at the edge of a greenhouse collects temperature, humidity, and wind speed data at 10-minute intervals, forming a standardized local micro-meteorological dataset with a time window from 08:00 to 12:00 on January 15, 2024. The forecast temperature, humidity, and wind speed data for this time period within the same geographic coordinate range are extracted using a public weather forecast service interface; the original forecast data interval is 1 hour. Cubic spline interpolation is applied to the forecast temperature series, inserting the corresponding forecast values ​​from 09:10 to 09:50 between the 09:00 and 10:00 temperature forecast values, and sampling is performed at the local sampling timestamp to obtain time nodes consistent with the local dataset. Taking humidity data as an example, piecewise quadratic spline interpolation is used to generate a smooth curve between the 09:00 forecast humidity of 70% and the 10:00 forecast humidity of 68%, and values ​​below 0% or above 100% due to forecast anomalies are removed. When interpolating wind speed data, a sudden gradient jump from 2 m / s to 8 m / s was detected in the original forecast from 09:00 to 10:00. A weighted moving average model was applied, with weights set to 0.6 for the current value and 0.4 for the previous value. After correction, the wind speed at the jump point was 5.6 m / s. After interpolation and resampling, the forecast temperature, humidity, and wind speed sequences were completely consistent with the timeline of the 10-minute resolution local dataset, forming a synchronized public weather forecast sequence. In validation, the time correspondence between this synchronized public weather forecast sequence and the local measured data reached a full match, ensuring complete consistency of the time base for subsequent short-delay comparisons and deviation vector calculations.

[0041] S1.3: The synchronous public weather forecast sequence is shifted backward using a preset short time delay offset, so that the shifted synchronous public weather forecast sequence matches the standardized local micro-meteorological dataset in terms of physical response time, so as to generate a time delay aligned public weather forecast sequence.

[0042] Using the synchronized public weather forecast sequence generated in step S1.2 as input, a preset short-delay time offset parameter is called to calculate the time-domain displacement of the sequence. This short-delay time offset is set based on the measured average thermal response delay of micrometeorological changes from inside the greenhouse to the outside, and is embedded as a time index adjustment factor in the displacement operation function during the calculation. The calculated displacement sequence is then subjected to sequence index increment operation, shifting the public weather forecast data at each time step backward to achieve a preliminary match between the time-domain data and physical delay characteristics. The synchronized public weather forecast sequence is then processed using a synchronization verification mechanism. Based on the timestamp sequence of the standardized local micrometeorological dataset, point-by-point time difference evaluation and matching degree scoring are performed to determine the degree of physical response time matching. For data segments with matching degree scores below a preset threshold, a secondary translation correction algorithm is called to perform fine-tuning displacement operations, eliminating local misalignments caused by short-delay estimation errors through interpolation compensation. The corrected, shifted public meteorological forecast sequence is strictly aligned in the time domain with the standardized local micrometeorological dataset to form a time-delay aligned public meteorological forecast sequence. A delay matching factor label is added to the output to ensure that subsequent differential operations can accurately identify the time-domain correspondence between meteorological changes and physical responses. Through short-delay displacement and delay matching verification, the synchronous public meteorological forecast sequence from the previous step is transformed into a time-delay aligned public meteorological forecast sequence that matches the local micrometeorological data in terms of physical response time, achieving the key time-domain fusion effect of connecting meteorological input with the actual thermal environment response of the greenhouse.

[0043] For example, for a group of multi-span glass greenhouses located at 39°N and operating in winter, when the average thermal response delay of the standardized local micrometeorological dataset is historically determined to be 18 minutes, the short-lag time offset parameter is set to 1080 seconds. The temperature, humidity, and wind speed data for each time step in the synchronous public weather forecast sequence are shifted backward by 1080 seconds. The difference between the shifted sequence and the timestamps of the local micrometeorological data is calculated and a matching score is given. Periods with scores below 0.85 undergo secondary correction. The secondary correction shift is calculated using the formula 1080 + Δ × 60 seconds, where Δ is the minute-level fine-tuning deviation. After correction, the weather forecast data and the standardized local micrometeorological data are fully aligned in the time domain, resulting in a time-lag aligned public weather forecast sequence. Verification shows that this sequence accurately reflects the 18-minute delay characteristic in the thermal response difference calculation, ensuring a significant improvement in the physical interpretability of the deviation in the subsequent element-by-element difference operation in S1.4, and greatly enhancing the rationality of the load forecast model's response to extreme temperature changes.

[0044] S1.4: Based on the standardized local micro-meteorological dataset and the time-delay aligned public meteorological forecast sequence, perform element-wise difference operation to calculate the temperature deviation component, humidity deviation component and wind speed deviation component, and combine the temperature deviation component, humidity deviation component and wind speed deviation component to generate an initial multidimensional meteorological forecast error vector.

[0045] Based on a standardized local micrometeorological dataset and a time-delay aligned public weather forecast sequence, temperature, humidity, and wind speed data components at corresponding time steps are extracted as dual input signals for element-wise difference operations. Linear difference calculations are performed on the temperature data components using the temperature deviation formula: in, These are measured temperature values ​​from the local micro-meteorological dataset. To align the predicted temperature values ​​in the public weather forecast series with time lag, the results are as follows: Characterizes the temperature deviation component. The humidity data component undergoes the same form of difference calculation, using the humidity deviation formula: in, For localized humidity measurements, The corresponding forecast humidity value, This represents the humidity deviation component. Differential calculations are performed on the wind speed data components, using the wind speed deviation formula: in, For local measured wind speed, To predict wind speed values, The wind speed deviation component is used as the input. The obtained temperature deviation, humidity deviation, and wind speed deviation components are combined according to their time index positions to form an initial multidimensional weather forecast error vector containing three ordered components. This vector serves as the direct input for subsequent sliding window aggregation processing. Through element-wise differencing and vector combination processing, the matching data from the previous step is transformed into an initial multidimensional index that can measure the magnitude and direction of the forecast error, achieving high-precision quantitative error output.

[0046] For example, five low-cost micro-weather stations were deployed at the edge of a greenhouse cluster. The standardized temperature sequences (18.5, 19.0, 19.4℃) were collected and differentially calculated with the time-lag aligned public weather forecast temperature sequences (17.8, 18.6, 19.1℃) to obtain temperature deviation components of 0.7, 0.4, and 0.3℃. The standardized humidity sequences (65, 60, 58%) were differentially calculated with the forecast humidity sequences (62, 61, 59%) to obtain a humidity deviation component of 3. 1, 1%. The collected standardized wind speed sequences (2.8, 3.1, 3.0 m / s) and the predicted wind speed sequences (2.5, 2.9, 2.8 m / s) were differencing to obtain wind speed deviation components of 0.3, 0.2, and 0.2 m / s. These three deviation components were combined according to time steps to obtain the initial multidimensional weather forecast error vector sequence: 0.7,3,0.3;0.4, 1, 0.2; 0.3, 1,0.2, This vector can significantly improve the model's ability to analyze the intensity and direction of forecast errors in subsequent confidence calculations.

[0047] S1.5: Perform sliding window aggregation and statistical feature encapsulation on multiple continuously generated initial multidimensional weather forecast error vectors at the hourly granularity, and mark the aggregated data stream as a time index to generate a forecast error vector sequence with temporal continuity.

[0048] Step S2: Based on the forecast error vector sequence, perform confidence quantification labeling to generate forecast confidence labels. Specifically, this includes: S2.1: Obtain the temperature deviation component, humidity deviation component, and wind speed deviation component in the forecast error vector sequence, and perform multivariate normalization and polar coordinate mapping processing on the temperature deviation component, humidity deviation component, and wind speed deviation component to generate a standardized error phase amplitude vector.

[0049] The temperature, humidity, and wind speed deviation components from the forecast error vector sequence are obtained. The multivariate data parsing module is then used to rigorously pair and assemble these three types of components into multidimensional vectors according to their time indices, forming the original error component matrix. A normalization operator is applied to this error component matrix, determining the normalization coefficient based on the historical statistical extreme values ​​and physically reasonable intervals of each component type. This maps the deviation values ​​of different dimensions to a unified numerical scale, eliminating scale differences in absolute values ​​for temperature, humidity, and wind speed. A polar coordinate mapping operator is then used to convert the normalized three-dimensional deviation vector into a combined representation of phase angle and amplitude. The amplitude is calculated using the following formula. in To normalize the temperature deviation component, To normalize the humidity deviation component, This represents the normalized wind speed deviation component. The phase angle is calculated using the following formula. This method characterizes the relative directional relationship between temperature and humidity deviations and can map wind speed deviation components to the radial coordinates of their amplitudes as needed. The amplitude and phase angle are combined and a time index is added to form a standardized error phase amplitude vector, which serves as the direct input for subsequent calculations of dynamic error fluctuation intensity. Through normalization and polar coordinate mapping, the forecast error vector sequence from the previous step is transformed into a standardized error phase amplitude vector with directional and magnitude-separable characteristics, providing the foundational data structure for subsequent reliable quantification labeling.

[0050] For example, considering the forecast error vector sequence of a greenhouse for 48 consecutive hours before a cold wave, the temperature deviation component is... The humidity deviation component is between 3.5 and 2.1. The wind speed deviation component is between 0.12 and 0.18. The normalization process uses coefficients corresponding to the extreme value intervals, dividing the temperature deviation by 5.0, the humidity deviation by 0.2, and the wind speed deviation by 2.0 to obtain the normalized temperature deviation. The range is Humidity deviation: 0.7 to 0.42 The range is 0.6 to 0.9, wind speed deviation The range is 0.9 to 0.75. Using the amplitude formula, within a certain hour... = 0.5 =0.8、 = At 0.4, the calculation yields The amplitude is approximately 1.02. Using the phase angle formula, the relationship between temperature and humidity deviation is obtained. The phase is approximately 2.13 radians. The standardized error phase amplitude vector output for this hour is (1.02, 2.13), which exhibits significant amplitude and direction characteristics in subsequent calculations of dynamic error fluctuation intensity, helping to maintain the fault tolerance and stability of the prediction model when a cold wave actually occurs.

[0051] S2.2: Based on the standardized error phase amplitude vector, the instantaneous rate of change and cumulative deviation of each component are calculated using a sliding time window. A weighted fusion operation is performed on the instantaneous rate of change and cumulative deviation to generate a dynamic error fluctuation intensity index.

[0052] S2.3: Based on the dynamic error fluctuation intensity index, a preset nonlinear confidence decay function is invoked for mapping transformation, and the output result of the nonlinear confidence decay function is discretized and graded to generate a basic confidence level index.

[0053] By obtaining the dynamic error fluctuation intensity index as an input parameter, the transient amplitude performance of the corresponding temperature, humidity and wind speed fluctuation characteristics in the time series can be clarified.

[0054] The preset nonlinear confidence decay function is invoked to map the dynamic error fluctuation intensity index to a confidence decay value. The function curve shape is used to compress high-amplitude fluctuations and remain sensitive to low-amplitude fluctuations, so as to ensure the nonlinear relationship between fluctuation strength and confidence change.

[0055] Based on the mapped confidence decay value, a discretization and grading process is performed. First, multiple confidence level threshold intervals are defined, and continuous confidence decay values ​​are divided into multiple level interval indices according to the intervals. The level category is represented by integer encoding.

[0056] Standardized encapsulation is performed on the discretized rank index to ensure that different rank categories have a uniform scale and searchability in the data structure.

[0057] By using nonlinear mapping and discretization hierarchical processing, the dynamic error fluctuation intensity index is transformed into a basic credibility level index, thereby achieving a quantifiable classification effect for the credibility of weather forecast errors.

[0058] S2.4: Combining the phase angle information of the standardized error phase amplitude vector with the basic confidence level index, perform directional symbol encoding and magnitude numerical encapsulation operations, and structurally concatenate the encoded symbol bits and encapsulated numerical bits to generate a forecast confidence label.

[0059] like Figure 2 As shown, step S3 involves constructing a set of thermal inertia parameters for greenhouses with different structural types, covering materials, and internal heat storage configurations, using the thermal response time constant, temperature hysteresis phase angle, and load attenuation coefficient obtained from offline calibration. Specifically, this includes: S3.1: Obtain static building archive data containing information on greenhouse structure type, covering material properties and internal heat storage configuration; perform multidimensional feature encoding and category mapping processing on the static building archive data to generate a greenhouse thermophysical structure fingerprint vector with a unique identifier.

[0060] It should be noted that the crop transpiration correction coefficient is used to characterize the degree of influence of crop transpiration on greenhouse latent heat load. Its value is obtained by calculating through a preset empirical model of transpiration rate based on crop type, growth stage and current humidity deviation, or by calibration through offline experiments. Under the condition of sudden increase in humidity and continued warm forecast, this coefficient is used to enhance the latent heat compensation term in load forecast.

[0061] Static building archive data containing information on greenhouse structure type, covering material properties, and internal heat storage configuration is obtained. Data is then integrated with the greenhouse management system archive, design drawing database, and project acceptance form through multi-source data interface calls to ensure data integrity and accuracy.

[0062] The merged static building archive data is processed by attribute domain separation, which extracts the structural type, covering material properties and heat storage configuration information into independent feature channels. In each channel, the data is categorized according to a predefined category code table, so that the structural type is converted into a discretized category index, the covering material properties are converted into material performance codes, and the internal heat storage configuration is converted into a heat storage capacity level identifier.

[0063] A multidimensional feature encoding algorithm is used to vectorize the above-mentioned discretized category index, material performance code and heat storage capacity level identifier, mapping various feature values ​​to a unified numerical space to form three feature sub-vectors.

[0064] A category mapping processing mechanism is introduced to combine feature sub-vectors into a set of composite structure vectors based on multi-dimensional vector splicing operation. A unique identifier generation algorithm is introduced during the merging process. This algorithm combines greenhouse number, geographical location, and structural parameters to generate a unique identifier to establish the unique identity of the greenhouse thermophysical structure.

[0065] By performing feature vector normalization, the composite structure vector is subjected to interval normalization and dimension consistency operations, so that each feature component is comparable within the numerical range, and finally a greenhouse thermophysical structure fingerprint vector with a unique identifier is formed.

[0066] Through the above-mentioned multidimensional feature encoding and category mapping processing, the static building archive data of the previous step is transformed into standardized and unique feature vectors that can be used for thermal inertia modeling, so as to achieve the expected technical effect of quantifying the diversity of greenhouse structures and thermophysical properties.

[0067] S3.2: Based on the greenhouse thermophysical structure fingerprint vector, call the preset transient thermal response simulation engine or historical measured dataset to perform dynamic thermal balance iterative calculations on the heat transfer process of the enclosure structure and the heat exchange process of the internal heat storage medium under different crop growth stages, so as to generate the original thermal dynamic trajectory sequence containing the temperature time domain response curve.

[0068] Based on the aforementioned greenhouse thermophysical structure fingerprint vector, a pre-built transient thermal response simulation engine is invoked to perform layered physical modeling of the heat transfer process of the enclosure structure under different crop growth stages. A multi-channel heat flow transfer path matrix consisting of radiative heat transfer, convective heat transfer, and conductive heat transfer is established, with each element value set according to the material's thermal conductivity, specific heat capacity, and surface radiation characteristics. State-space modeling is performed on the heat exchange process of the internal heat storage medium, using medium temperature, heat flux, and mass heat storage as state variables to form a set of state equations. The coefficients of the state equations are determined based on the characteristic parameters of the heat storage medium and the crop canopy transpiration parameters. A dynamic thermal balance iterative algorithm is used to simultaneously solve the heat transfer equations of the enclosure structure and the heat exchange equations of the heat storage medium within discrete time steps, ensuring that the heat flow output at each time step is consistent with the updated temperature field value inside the greenhouse. Numerical integration is used to calculate the temperature change at each time step, and the output temperature sequence is used to construct a temperature time-domain response curve according to the time index. The heat input and external meteorological boundary conditions are corrected using different diurnal and nighttime boundary functions according to the crop growth stage. To ensure physical consistency, energy conservation verification is performed on the iterative calculation results, the difference between input heat and output heat is limited to a preset tolerance threshold range, and the temperature time-domain response curve that meets the requirements is encapsulated into the original thermal dynamic trajectory sequence.

[0069] Through the above dynamic thermal balance iterative calculation and energy conservation verification process, the greenhouse thermophysical structure fingerprint vector from the previous step is transformed into original thermal dynamic trajectory data containing the temperature change law over time, realizing offline simulation output of greenhouse thermal inertia characteristics under different structures and different crop stages.

[0070] For example, in a greenhouse with a double-layered inflatable membrane structure, a covering material with a thermal conductivity of 0.03 W / (m·K), and an internal water wall heat storage system, the plants are in the flowering stage. The transient thermal response simulation engine sets the external boundary conditions as follows: daytime temperature 25℃, nighttime temperature 15℃, relative humidity 60% and 80%, respectively, and wind speed 0.8 m / s during the day and 0.3 m / s at night. The simulation engine uses a thermal conductivity of 0.03 W / (m·K), a specific heat capacity of 1250 J / (kg·K), and a surface emissivity of 0.85 for the heat transfer process of the enclosure structure; a water wall heat storage medium density of 1000 kg / m³ and a specific heat capacity of 4200 J / (kg·K); and a daytime transpiration gain coefficient of 1.2 for the crop canopy. In the iterative calculation, the discrete time step of the heat balance equation is 5 minutes. Numerical integration yields a temperature time-domain response curve with a peak lag of approximately 1 hour within the diurnal cycle. The curve shape is verified by energy conservation, and the difference between input and output heat remains within the allowable 2500 J. This temperature time-domain response curve is encapsulated as an original thermal dynamic trajectory sequence. In subsequent system identification algorithms, a thermal response time constant of 2.5 hours, a temperature lag phase angle of 15°, and a load attenuation coefficient of 0.78 can be extracted, significantly improving the adaptability and prediction accuracy of the thermal inertia parameter set for flowering conditions.

[0071] S3.3: The system identification algorithm is used to perform parameter extraction processing on the original thermal dynamic trajectory sequence. The thermal response time constant, which characterizes the speed of heat transfer, the temperature lag phase angle, which characterizes the degree of temperature peak lag, and the load attenuation coefficient, which characterizes the load fluctuation attenuation amplitude, are separated and quantified from the temperature time domain response curve to generate a discretized thermal inertia characteristic parameter set.

[0072] Based on the temperature time-domain response curves in the original thermal dynamic trajectory sequence and the fingerprint vector of the greenhouse thermophysical structure, the system identification algorithm is called to perform parameter extraction processing. The curve data is first preprocessed by detrending and denoising to eliminate background drift and high-frequency noise interference.

[0073] An exponential fitting function is applied to the denoised temperature time-domain response curve to establish a first-order or multi-order heat transfer model, in order to separate the exponential decay factor corresponding to the thermal response time constant.

[0074] The phase spectrum analysis method is used to perform Fourier transform on the temperature time-domain response curve and the excitation signal, extract the phase difference data and convert it into temperature hysteresis phase angle value.

[0075] The load fluctuation attenuation amplitude is calculated using energy ratio analysis and amplitude envelope fitting methods, and the results are converted into load attenuation coefficients.

[0076] The aforementioned thermal response time constant, temperature hysteresis phase angle, and load attenuation coefficient are uniformly encapsulated into a discretized set of thermal inertia characteristic parameters.

[0077] The thermal response time constant is calculated using the following formula: in, The thermal response time constant, This is the exponential decay fitting factor.

[0078] The formula for calculating the temperature hysteresis phase angle is as follows: in, This is the temperature hysteresis phase angle. Peak latency For the incentive cycle.

[0079] The load attenuation coefficient is calculated using the following formula: in, This is the load attenuation coefficient. The amplitude after attenuation. This is the initial amplitude.

[0080] By using the system identification algorithm and the above formula processing method, the original thermal dynamic trajectory sequence of the previous step is transformed into a set of discretized thermal inertial feature parameters that can be quantified and called, thereby realizing the structured input of thermophysical properties into the prediction model.

[0081] S3.4: Perform a three-dimensional association binding operation on the discretized thermal inertia feature parameter group, the greenhouse thermophysical structure fingerprint vector, and the corresponding typical crop growth stage labels. Perform normalization and outlier smoothing on the bound data to generate a standardized thermal inertia reference parameter library.

[0082] Based on the discretized thermal inertial characteristic parameter set, greenhouse thermophysical structure fingerprint vector, and typical crop growth stage labels, a three-dimensional association binding operation is performed to achieve close coupling between structure, crop, and climate elements.

[0083] The multidimensional normalization transformation function is called on the data set after three-dimensional association and binding to perform dimensional unification processing on numerical features including thermal response time constant, temperature hysteresis phase angle and load attenuation coefficient, so as to eliminate the order of magnitude difference between different features.

[0084] A standardization transformation is performed on the normalized dataset. A feature space with zero mean and unit variance is established by mean centering and standard deviation scaling to ensure the consistency of numerical stability of the parameters used in subsequent model calls.

[0085] Outlier detection is performed on the standardized dataset. Sliding window statistics and Mahalanobis distance threshold are used to identify parameter points that deviate from the normal distribution. A smooth interpolation algorithm is then called to fill in the outliers between adjacent time series points, maintaining the continuity and availability of the parameter library.

[0086] The standardized dataset, after outlier smoothing, is encapsulated into a thermal inertial reference parameter library to form a stable and physically consistent parameter storage unit that can be called by subsequent fast retrieval mechanisms.

[0087] Through the above processing method, the discretized thermal inertia characteristic parameter set of the previous step is transformed into a thermal inertia benchmark parameter library with unified dimensions, stable numerical distribution and anomaly suppression capability, so as to realize the high adaptability and high robustness of the load prediction model in calling parameters under different greenhouse structures and climate conditions.

[0088] S3.5: Based on the standardized thermal inertia reference parameter library, establish a fast retrieval mapping mechanism from real-time operating conditions to specific parameter values, and perform index encapsulation processing on the output results of the fast retrieval mapping mechanism to generate a thermal inertia parameter set.

[0089] Based on the standardized thermal inertia reference parameter library's indexing system, the real-time operating conditions of the current greenhouse are obtained, including structural type, covering material type, crop growth stage, and internal heat storage configuration. Multi-dimensional feature encoding is performed on these real-time operating conditions to generate a retrieval key matching the reference parameter library. This retrieval key is input into a pre-set fast retrieval mapping mechanism, which uses a combination of hash mapping and hierarchical indexing to sequentially pass the features of each dimension of the operating conditions to the hierarchical search node, locating the corresponding thermal inertia feature parameter group. The retrieved thermal inertia feature parameter group undergoes index encapsulation processing. During encapsulation, a unique identifier is used to bind three core parameters: thermal response time constant, temperature lag phase angle, and load attenuation coefficient, forming structured parameter entries. The parameter entries are double-indexed using a unique identifier and a timestamp to ensure that subsequent model calls can retrieve and obtain the corresponding physical parameter values ​​within milliseconds. The double-indexed thermal inertia parameter entries are stored in a high-performance cache so that the load prediction model can directly call them when performing dual-channel feature fusion, achieving rapid mapping from operating conditions to specific parameter values. Through the above processing method, the standardized thermal inertia reference parameter library of the previous step is transformed into a thermal inertia parameter set with real-time retrieval capability, so as to realize the rapid parameter adaptation and physical consistency guarantee of the load prediction model under different production conditions.

[0090] For example, in a certain facility horticulture base, the current greenhouse operating conditions are a multi-span glass greenhouse, Low-E coated covering, and internal water wall heat storage, with the crops in the flowering stage. Real-time operating condition feature encoding generates a retrieval key vector [3,5,2,4], where each element represents the category number of the structure type, covering material, heat storage configuration, and crop stage, respectively. A fast retrieval mapping mechanism uses a hierarchical index. The retrieval key is passed to the first-level structure type index node to locate the multi-span glass greenhouse cluster, then to the second-level covering material node to locate the Low-E coated covering category, the third-level heat storage configuration node to locate the water wall cluster, and the fourth-level crop stage node to locate the flowering stage cluster. Finally, a hash mapping is used to locate the entry with the unique identifier ID_327 in the parameter library. This entry contains a thermal response time constant of 3.6 hours, a temperature lag phase angle of 25 degrees, and a load attenuation coefficient of 0.42. During the encapsulation process, a unique identifier ID_327 is bound to the three parameters mentioned above, along with the current timestamp 1693826400, forming a complete structured parameter entry, which is then stored in the cache. When the load forecasting model is invoked, this entry is quickly retrieved and combined with the credibility tag of the concurrent weather forecast to participate in the dual-channel feature fusion calculation. This enables the model to accurately correct the thermal inertia of the current operating conditions. Subsequent comparisons show that the average absolute deviation between the forecast results and the measured values ​​is significantly reduced, meeting the technical goal of high-precision forecasting.

[0091] like Figure 3 As shown, step S4 involves acquiring conventional weather forecast elements, using the time series features of these elements as the primary input, and the forecast confidence label and the corresponding thermal inertia parameter set index as auxiliary inputs. Feature fusion processing is then performed to generate a fused feature vector. Specifically, this includes: S4.1: Obtain conventional weather forecast elements and their time series characteristic data, perform sliding window statistics calculation and frequency domain transformation processing on the conventional weather forecast elements and their time series characteristic data, and generate the main channel time series characteristic matrix.

[0092] It should be noted that the conventional meteorological forecast elements include at least forecast temperature, forecast humidity, and forecast wind speed, which are key meteorological parameters extracted from public meteorological forecast data that have the most direct impact on greenhouse heat load.

[0093] When acquiring conventional meteorological forecast elements and their time-series characteristic data, a pre-defined meteorological data interface is invoked to extract elements such as forecast temperature, forecast humidity, and forecast wind speed, and their timestamp information and geospatial index are synchronized to form the original main channel element sequence. A sliding time window segmentation operation is performed on the forecast temperature, humidity, and wind speed sequences, calculating statistics such as mean, maximum, minimum, and standard deviation within each window to characterize the variation pattern of meteorological elements within a local time period. Based on the aforementioned window statistical vectors, a fast Fourier transform is performed on each meteorological element to generate a frequency domain spectral coefficient matrix, and the energy distribution characteristics of periodically changing components are captured using power spectral density estimation. For spectral sequences with multiple frequency peaks in the frequency domain, bandpass filtering separation processing is performed to extract the amplitude and phase parameters of the corresponding periodic components, and the characteristics of each periodic component are arranged according to the time window order to form a comprehensive feature representation in both the time and frequency domains. The comprehensive feature representation is normalized and dimensionally unified to ensure that the time-domain statistical features and frequency-domain spectral features of temperature, humidity, and wind speed can be directly compared and combined on the same numerical scale, thus forming a main channel time-series feature matrix characterizing meteorological change trends. Through the above processing method, the results of the previous step are transformed into a main channel time-series feature matrix containing multi-time-scale statistics and periodic spectral features, realizing a quantitative characterization and structured expression of the changing trends of meteorological elements.

[0094] For example, in a greenhouse cluster management scenario, forecasted temperature, humidity, and wind speed sequences from a public weather forecast service are obtained, with a time resolution of 10 minutes and aligned to local standard time. A sliding window length of 60 minutes and a window step size of 10 minutes are set. For each window, the mean, standard deviation, maximum, and minimum values ​​of the temperature sequence are calculated; the mean, standard deviation, and range of the humidity sequence are calculated; and the mean and peak velocity of the wind speed sequence are calculated, forming a statistical feature vector of length 6. A Fast Fourier Transform (FFT) is performed on the temperature, humidity, and wind speed sequences for each window, with an FFT length of 64 points, extracting the amplitude and phase of the frequency component corresponding to the maximum power spectral density. Bandpass filters are used to retain components with periods of 2 hours and 4 hours, respectively, obtaining their amplitude and phase features, and recording them in chronological order. The time-domain statistical features are concatenated with the frequency-domain amplitude and phase features to form a comprehensive feature vector of length 18, and min-max normalization is performed on all vectors to unify their value range between 0 and 1. The final output main channel time series feature matrix contains an 18-dimensional feature vector sequence of N time windows, where N is determined by the available data length and window setting. In the subsequent S4.2 step, this matrix will interact with the auxiliary channel state vectors across modal features, significantly improving the interpretability and prediction robustness of meteorological change trends.

[0095] S4.2: Based on the time index of the main channel time series feature matrix, the corresponding forecast confidence label and thermal inertia parameter set index are extracted by calling the preset mapping retrieval mechanism. One-hot encoding mapping and numerical normalization processing are performed on the forecast confidence label and thermal inertia parameter set index to generate the auxiliary channel state vector.

[0096] Based on the time index of the main channel time series feature matrix, a preset mapping retrieval mechanism is invoked to extract the prediction confidence label and thermal inertial parameter set index for the corresponding time from the thermal inertial reference parameter library.

[0097] The extracted forecast confidence labels are deconstructed into directional sign bits and magnitude value bits. One-hot coding mapping is performed on the directional sign bits to map each possible direction (e.g., higher temperature, lower temperature, higher humidity, lower humidity, higher wind speed, lower wind speed) into a multidimensional sparse coding format of unit basis vectors.

[0098] The magnitude values ​​are normalized according to a preset physical dimension reference table to unify the error magnitudes corresponding to different dimensions (such as degrees Celsius, relative humidity percentage, and wind speed in meters per second) within the same numerical range for subsequent model calculations.

[0099] The thermal inertia parameter set is indexed and parsed into physical characteristic components such as thermal response time constant, temperature lag phase angle, and load attenuation coefficient. Interval normalization mapping is performed on the thermal response time constant, period normalization is performed on the temperature lag phase angle, and proportional normalization is performed on the load attenuation coefficient, so that each physical characteristic is kept within a uniform characteristic value range.

[0100] The normalized error direction magnitude information and the normalized thermophysical property information are concatenated and structured according to time index to construct an auxiliary channel state vector containing multimodal information, ensuring that it has strict synchronization with the main channel temporal feature matrix in the time domain.

[0101] By using the aforementioned unique thermal encoding and numerical normalization processing, the retrieval results from the previous step are transformed into auxiliary channel state vectors that can directly participate in cross-modal feature interaction, thereby accurately injecting the magnitude of meteorological error direction and the characteristics of greenhouse thermophysical properties into the feature fusion stage of the prediction model.

[0102] S4.3: Using an attention mechanism weight allocation algorithm, cross-modal feature interaction operations are performed on the main channel time series feature matrix and the auxiliary channel state vector. The feature weights of each time step of the main channel are adjusted according to the error fluctuation intensity in the forecast confidence label to generate a weighted meteorological feature tensor.

[0103] The synchronization time index of the main channel temporal feature matrix and the auxiliary channel state vector is obtained. The error fluctuation intensity value in the forecast confidence label is used as a weight adjustment parameter and input to the scoring function module of the attention mechanism weight allocation algorithm. The meteorological time step feature importance assessment operation is performed on the scoring function module. The dot product correlation calculation is performed between the feature vector of each time step of the main channel and the state vector of the auxiliary channel to generate a cross-modal matching score matrix. Normalized weight calculation is performed on the cross-modal matching score matrix. The score matrix is ​​transformed into a probability distribution through the softmax normalization function to ensure that the sum of the weights at each time step is 1. At the same time, the error fluctuation intensity is retained as a dynamic input for the softmax temperature factor, thereby suppressing or enhancing the feature weights during periods of extreme fluctuations.

[0104] The main channel feature matrix is ​​weighted and summed using the probability distribution weights. Meteorological forecast features at different time steps are reconstructed according to their importance under the current error state to generate a weighted meteorological feature tensor corrected for uncertainty. Dimensionality consistency checks and numerical range constraints are applied to the weighted meteorological feature tensor to ensure that the output tensor meets the physical dimension and structural consistency requirements for subsequent feature concatenation. Through cross-modal feature interaction computation using an attention mechanism, the main and auxiliary channel data from the previous step are transformed into a weighted meteorological feature tensor that can dynamically adapt to fluctuations in meteorological forecast errors, thereby enhancing the robustness of load forecast feature input under extreme weather conditions.

[0105] For example, considering meteorological forecast data for a greenhouse cluster, the main channel time-series feature matrix contains statistical features of temperature, humidity, and wind speed at 24 time steps, each time step being a 12-dimensional feature vector. The auxiliary channel state vector consists of the error direction magnitude label and thermal inertia parameters at the corresponding time point, totaling 4 dimensions. The error fluctuation intensity reaches high values ​​of 2.8 and 3.2 at time steps 8 and 15, respectively, with temperature factors τ of 0.75 and 0.70 after normalization mapping. The main channel feature matrix... With auxiliary channel state vector Performing a dot product operation yields a 24×1 matching score matrix. After softmax temperature adjustment normalization, the weights at time steps 8 and 15 are reduced to 0.05 and 0.04, respectively, while the weights at other time steps increase between 0.03 and 0.07. This weight vector is then used to calculate the weighted sum of the original main channel feature matrix, generating a 24×12 weighted meteorological feature tensor with values ​​normalized to 0-1. This tensor significantly reduces the weight of instantaneous low-temperature features during extreme temperature fluctuations, preventing the model from overreacting to short-term abnormal predictions. This verifies that the output load prediction bias is significantly reduced compared to the original model, and the fluctuation amplitude of the prediction curve tends to stabilize under physical consistency constraints.

[0106] S4.4: Based on the weighted meteorological feature tensor and the thermophysical information in the auxiliary channel state vector, perform feature splicing and nonlinear activation function mapping processing to bind the meteorological error propagation path and the greenhouse thermal inertial response delay characteristics in the feature space to generate an intermediate feature vector.

[0107] The weighted meteorological feature tensor, after uncertainty correction, and the thermophysical attribute information in the auxiliary channel state vector are used as inputs. Feature dimension matching operations are performed on the two types of data to ensure that the index correspondence of the tensors does not shift during the splicing process. Feature splicing processing is then applied to the matched tensor data, merging the time-series meteorological elements of the weighted meteorological feature tensor with the static building thermal inertia parameters of the thermophysical attribute vector in spatial dimension to form a composite feature matrix containing both dynamic and static information. A nonlinear activation function mapping mechanism is applied to this composite feature matrix to interact with the error propagation path and thermal inertia response delay characteristics in the greenhouse meteorological input in the feature space, transforming the high-dimensional nonlinear relationship using the activation function. A hybrid activation strategy of hyperbolic tangent and modified linear unit is employed to perform amplitude compression and sparsification processing on different ranges of the composite feature matrix to extract feature components highly sensitive to load forecasting and suppress noise features. The activated and mapped feature matrix is ​​then coupled and bound. A parameterized binding operator establishes a one-to-one mapping relationship between the time-domain shift of the meteorological error vector and the phase-domain delay characteristics of the thermal inertia coefficient, thereby generating a preliminary fused intermediate feature vector. By using feature concatenation and nonlinear mapping, the results of the previous step are transformed into data that includes the dynamic state of weather forecast error and the physical delay of greenhouse thermal inertia, which can be used as input for subsequent gating modules. This enhances the physical consistency and interpretability of the input features of the load forecasting model.

[0108] For example, in the actual operating conditions of a group of greenhouses, the weighted meteorological feature tensor after uncertainty correction includes three types of time-series features: temperature, humidity, and wind speed. Each feature is sampled once per hour, for a total of 24 time steps, corresponding to a tensor dimension of 24×3. The auxiliary channel state vector consists of a thermal response time constant τ = 5.2 hours, a temperature lag phase angle φ = 18 degrees, and a load attenuation coefficient k = 0.85. After performing feature dimension matching, the two inputs are concatenated in the column direction to obtain a composite feature matrix of dimension 24×6. The hyperbolic tangent activation function is then applied to each element of this matrix. With modified linear unit Mixed mapping based on different index ranges, the mapped values ​​are in [ The interval [1,1] is compressed or its non-negative sparse distribution is preserved. A binding operator is used to combine the temperature deviation component of the meteorological error vector with... Humidity deviation component and Wind speed deviation component and The mapping relationship is established, and its binding formula is expressed as follows: ,in It is a ternary coupling mapping matrix. This is a composite feature vector. After performing this binding operation, a preliminary fused intermediate feature vector is obtained. Verification results show that when this intermediate feature vector is used in the subsequent error-inertia coupling gating module, the response curve of the load prediction model under cold wave and rainstorm conditions is significantly improved in stability, avoiding overly aggressive adjustment actions caused by sudden low temperatures or humidity increases.

[0109] S4.5: Perform dimensionality reduction projection and orthogonalization cleaning on the intermediate feature vector to remove redundant feature components and unify feature dimensions, generating a fused feature vector.

[0110] The intermediate feature vectors obtained from the initial fusion are used as input objects. Dimensionality reduction projection processing based on the feature covariance matrix is ​​performed on the input feature vectors to compress redundant feature dimensions and retain the main information.

[0111] Calculate the feature covariance matrix and solve for its eigenvalues ​​and eigenvectors. Based on the cumulative contribution rate threshold, select the eigenvectors corresponding to the top K largest eigenvalues ​​to construct a dimension reduction transformation matrix, mapping the original eigenvectors to a lower-dimensional space, as shown in the following formula: in The feature vectors after dimensionality reduction. It is the projection matrix composed of the first K eigenvectors. This is the initial eigenvector matrix.

[0112] The dimensionality-reduced feature data is cleaned by orthogonalization. The Gram-Schmidt orthogonalization algorithm is used to eliminate the linear correlation between different feature components, ensuring that each component of the feature vector is spatially independent.

[0113] The orthogonalized eigencomponents are normalized to unify their dimensions. Z-score standardization is used to convert the eigenvalues ​​of different physical quantity ranges into a standard scale with a mean of zero and a standard deviation of one. The normalization formula is as follows: in These are the original eigenvalues. The mean of this feature. The standard deviation of this feature. These are the normalized eigenvalues.

[0114] Through the aforementioned methods of dimensionality reduction projection, orthogonalization cleaning, and dimensional unification, the intermediate feature vectors that are initially fused are transformed into fused feature vectors with high interpretability and physical consistency, enabling direct and efficient input to the error inertial coupling gating module.

[0115] For example, in a greenhouse meteorological error compensation modeling scenario, the initial fused intermediate feature vector has a dimension of 48, including 24 dimensions of meteorological trend features, 12 dimensions of thermal inertia features, and 12 dimensions of crop physiological features. The top 10 principal components with a cumulative contribution rate of 0.92 are selected to form a projection matrix. After projection transformation, the dimension is reduced to 10, preserving the main information and significantly improving vector processing efficiency. Gram-Schmidt orthogonalization is performed on the dimensionality-reduced data to eliminate linear correlations where the correlation coefficient of each component exceeds 0.05. The orthogonalized feature components are spatially independent. Z-score standardization is performed on the orthogonalized features. The mean μ and standard deviation σ are set based on the statistical values ​​of the feature in the full training dataset. For example, the mean of a certain temperature change rate feature is 0.8, the standard deviation is 0.12, and the normalization result is calculated to be 1.25. The fused feature vector output by the above processing triggers matching logic in the subsequent error inertia coupling gating module, achieving a significant improvement in load forecasting under extreme weather conditions and ensuring consistency with physical constraints.

[0116] Step S5: Based on the abrupt change intensity of the prediction error vector in the fused feature vector, dynamically adjust the thermal inertia compensation weights to generate a control signal. Specifically, this includes: S5.1: Obtain the prediction error vector component in the fused feature vector, and perform sliding window second-order difference operation and gradient magnitude calculation on the prediction error vector component to generate a scalar sequence of error mutation intensity.

[0117] Obtain the forecast error vector components in the fused feature vector, and explicitly extract the numerical sequences of temperature deviation components, humidity deviation components, and wind speed deviation components as the calculation objects.

[0118] The extracted deviation component sequences are divided into sliding windows based on a unified time index. The sequence length within each window is set to cover the time domain range of meteorological change characteristics, ensuring that the data within the window can reflect transient meteorological fluctuations.

[0119] Within each sliding window, a second-order difference operation is performed on the sequence. By calculating the second-order increment of the sequence at adjacent time steps, the acceleration change characteristics of the weather forecast deviation over time are captured. The second-order difference formula is expressed as: in, These are the deviation component values. For time indexing.

[0120] The gradient modulus of the second-order difference operation results is calculated in each dimension of the deviation component to quantify the instantaneous synthetic abrupt change intensity of multidimensional meteorological deviations. The formula for calculating the gradient modulus is as follows: in, Indicates the second-order difference. , , These represent the deviation components of temperature, humidity, and wind speed, respectively.

[0121] A time series reconstruction operation is performed on the gradient magnitude sequence to ensure that the calculation results correspond one-to-one with the original deviation sequence in the time dimension, and to ensure that the subsequent mutation intensity index can be directly called by the dynamic adjustment module.

[0122] Through the above-mentioned sliding window second-order difference and gradient magnitude calculation processing, the error components in the fused feature vector of the previous step are transformed into a scalar sequence of error mutation intensity that characterizes the transient severity of weather forecast deviation, thereby realizing the quantitative identification capability of mutation events.

[0123] S5.2: Based on the error mutation intensity scalar sequence, call the preset nonlinear mapping function library to perform adaptive threshold segmentation processing, map the continuously changing error mutation intensity scalar sequence into a discretized meteorological uncertainty level index, and generate a classification status identifier.

[0124] Based on the scalar sequence of error mutation intensity generated by S5.1, the stepwise piecewise Sigmoid, symmetric hyperbolic tangent, and exponential decay function modules from the pre-built nonlinear mapping function library are called. At each time step, the error mutation intensity is input into the corresponding function for amplitude transformation, forming a multi-channel nonlinear response matrix. Row vector normalization is performed on the multi-channel nonlinear response matrix to ensure that the output values ​​of different function channels are within a unified normalization range. The local peak position and gradient change trend across time steps of each channel output are calculated to form a composite feature set that can characterize the mutation amplitude and mutation persistence. According to the correlation coefficient between the peak position and gradient change trend in the composite feature set, an adaptive threshold segmentation algorithm is called to dynamically adjust the lower and upper limits of the mutation intensity threshold, realizing real-time demarcation between normal fluctuations and extreme mutations. The threshold segmentation result is encoded into a discretized meteorological uncertainty level index. The index value is divided into multi-level classification intervals according to the degree of mutation. It is converted into a classification status identifier through an index mapping table for subsequent S5.3 to call the classification status identifier to match the corresponding thermal inertia parameters.

[0125] By using a multi-channel response of a nonlinear mapping function and an adaptive threshold segmentation method, the continuous error mutation intensity scalar sequence generated in the previous step is transformed into a discrete meteorological uncertainty level index that can be directly used for subsequent thermal inertia weight adjustment, thereby enabling rapid classification and determination of different mutation modes and improving response accuracy under extreme weather conditions.

[0126] S5.3: Extract the thermal response time constant and temperature hysteresis phase angle corresponding to the classification state identifier from the thermal inertia parameter set, construct a dynamic weight allocation model using the hyperbolic tangent activation function, and perform nonlinear weighted fusion operation on the thermal response time constant and temperature hysteresis phase angle to generate initial thermal inertia compensation weight coefficients.

[0127] Based on the classification status identifier, an index mapping relationship is established between the meteorological uncertainty level index and the structure-crop-climate ternary coupled thermal inertial parameter library, and the thermal response time constant and temperature lag phase angle corresponding to the current level are extracted to form a feature tuple.

[0128] The thermal response time constant and temperature hysteresis phase angle are standardized in terms of units and dimensions to ensure that the contribution ratio of each parameter is controllable in subsequent weight calculations.

[0129] A hyperbolic tangent activation function is constructed as the core of the nonlinear response of the dynamic weight allocation model. The function's form is set as follows: in The normalized thermal response time constant, To normalize the temperature hysteresis phase angle, , This is the weighting ratio coefficient.

[0130] By performing a nonlinear weighted fusion operation on the feature binary tuple using the hyperbolic tangent activation function described above, an initial compensation weight coefficient vector reflecting the combined thermal inertia characteristics is obtained.

[0131] Boundary truncation and signal smoothing are performed on the generated initial compensation weight coefficient vector to suppress weight oversaturation and oscillation problems caused by abrupt changes in level.

[0132] By using dynamic parameter mapping and nonlinear fusion, the classification state identifiers from the previous step are transformed into initial thermal inertia compensation weight coefficients that are adaptable to the current physical conditions, thereby enabling adaptive adjustment of the model's thermal hysteresis response intensity.

[0133] S5.4: Using the initial thermal inertia compensation weight coefficient, combined with the crop transpiration correction coefficient in the fused feature vector, the proportional-integral-derivative adjustment algorithm is executed to perform phase lag compensation correction, and the corrected weight coefficient is subjected to normalization constraint processing to generate a standardized thermal inertia dynamic adjustment factor.

[0134] The input group for composite regulation is established by using the initial thermal inertia compensation weight coefficient and the crop transpiration correction coefficient in the fused feature vector. The signal analysis module performs data alignment of the two at the same resolution so that they can enter the proportional-integral-derivative (PID) regulation stage.

[0135] A PID controller is used to calculate the proportional, integral, and derivative terms respectively. The proportional term is directly multiplied by the input coefficient based on the transient thermal inertia deviation. The integral term forms a cumulative correction factor by sliding and accumulating the thermal inertia deviation. The derivative term uses the time gradient of the deviation to predict and suppress the lag trend. The three results are weighted and synthesized into a phase lag compensation value.

[0136] The phase lag compensation value of the PID control output is used to additively adjust the initial thermal inertia compensation weight coefficient to form a corrected thermal inertia weight vector. In the adjustment process, the crop transpiration correction coefficient is introduced as a correction gain to reflect the change in latent heat contribution.

[0137] The modified thermal inertia weight vector is subjected to normalization constraint processing, and each element is linearly scaled to a set standardized interval according to the physically feasible range, using the normalization formula. in These are the weighting coefficients. and These represent the lower and upper limits allowed by physics, achieving unified dimensions and physical interpretability.

[0138] By using PID phase lag compensation and normalization constraints, the dynamic weight correction from the previous step is transformed into a standardized thermal inertia dynamic adjustment factor, thereby achieving physical interpretability and enabling the direct driving of the adjustment signal output of the prediction model.

[0139] For example, under the conditions of a glass-enclosed greenhouse structure, double-layer inflatable film covering, and phase change material plate heat storage, the initial thermal inertia compensation weight coefficient is set to 0.65, the crop is tomato fruiting stage, and the transpiration enhancement factor is 1.12. Through input alignment, the PID proportional term coefficient is 0.15, the integral term cumulative value is 0.08, and the derivative term is... The three terms are combined to obtain a phase lag compensation value of 0.21. The initial weight coefficients are adjusted using the compensation value to obtain a correction value of 0.86, which is then multiplied by the evaporation enhancement factor to obtain 0.9632. The normalization interval is set to [0.5, 1.0].

[0140] Substituting the correction value into the formula, the calculation result is 0.9264, forming the standardized thermal inertia dynamic adjustment factor. Validation results show that in the extreme condition where the simulated temperature forecast drops by 5℃ but the actual temperature drop is only 1.5℃, this adjustment factor drives the prediction model to suppress the overestimation of instantaneous load, significantly reducing the peak value of the load curve and bringing it closer to the boundary of the physically feasible region, thus achieving effective compensation for extreme weather errors.

[0141] S5.5: Based on the standardized thermal inertia dynamic adjustment factor and the meteorological uncertainty level index, perform logic gating coding and pulse width modulation signal synthesis processing to convert the numerical adjustment factor into a binary control flow that can drive the switching state of neurons inside the load prediction model, and generate the control signal.

[0142] The input conditions for the gated driving logic are constructed based on the standardized thermal inertia dynamic adjustment factor and the meteorological uncertainty level index, and the numerical factors and hierarchical index containing weight coefficients are selected as the processing objects.

[0143] Binary logic encoding is performed on the meteorological uncertainty level index, mapping the level index into a logical state sequence according to a preset threshold, and the control state corresponding to different levels is represented by a fixed-length bit sequence through encoding rules.

[0144] The amplitude quantization process is performed on the standardized thermal inertia dynamic adjustment factor. The continuous factor is divided into discrete amplitude levels according to the resolution, and each level is mapped to the corresponding pulse width duty cycle.

[0145] The logic state sequence is associated and matched with the pulse width increment sequence, and a structured gating state table is generated under the matching relationship to guide the subsequent control pulse synthesis.

[0146] The pulse width modulation (PWM) signal synthesis algorithm is invoked to couple the logic state and duty cycle information in the gating state table into a binary control flow that can drive the switching state of the model neuron. By modulating the high and low level time ratio in each control cycle, physical-level driving of the neuron state is achieved.

[0147] By using logic gating encoding and PWM signal synthesis processing, the standardized thermal inertia dynamic adjustment factor and meteorological uncertainty level index from the previous step are transformed into the final control signal, enabling precise switching and dynamic state adjustment of neurons within the load forecasting model.

[0148] For example, in a cluster of greenhouse facilities for agriculture, the amplitude range of the standardized thermal inertia dynamic adjustment factor is set to 0 to 1, with a resolution of 0.05, and the meteorological uncertainty level index is divided into 0 to 7 levels. States with an index below 3 are encoded as logic "00", 3 to 5 as logic "01", and greater than 5 as logic "10". The thermal inertia adjustment factor of 0.00 to 0.25 is mapped to a duty cycle of 20%, 0.30 to 0.50 to a duty cycle of 40%, 0.55 to 0.75 to a duty cycle of 60%, and 0.80 to 1.00 to a duty cycle of 80%. After matching the logic state with the duty cycle, the following gating state table is generated: when the logic is "00" and the duty cycle is 20%, a low-power response is driven; when the logic is "01" and the duty cycle is 40%, a medium-power response is driven; and when the logic is "10" and the duty cycle is 80%, a high-power response is driven. In PWM signal generation, each control cycle is set to 10ms. With a 20% duty cycle, the high level lasts for 2ms and the low level for 8ms; with an 80% duty cycle, the high level lasts for 8ms and the low level for 2ms. This PWM signal is input into the internal control interface of the load forecasting model. The model's neuron state is dynamically adjusted according to the ratio of high to low level durations, significantly improving response speed and physical delay characteristics. The verification results of the above embodiments show that when the intensity of sudden changes in weather forecast errors is high and greenhouse thermal inertia is large, high duty cycle driving can significantly improve the model's load forecasting accuracy and stability under extreme weather conditions.

[0149] Step S6: The control signal is used to perform feedforward offset correction processing on the load change trend. When a sudden drop in forecast temperature is detected but local measured temperature drop is slow, the thermal inertia delay response mechanism is activated; or when the forecast is consistently warmer than normal and local humidity rises sharply, the crop transpiration correction coefficient is invoked to generate a predicted heat load value under the constraints of the physically feasible region. Specifically, this includes: S6.1: Obtain the binary control flow and standardized thermal inertia dynamic adjustment factor in the control signal, perform logic state parsing processing on the binary control flow to identify transient operating condition characteristics of a sudden drop in temperature forecast and slow local measured temperature drop, and call the preset thermal inertia delay response algorithm based on the identified transient operating condition characteristics to perform phase lag compensation operation on the standardized thermal inertia dynamic adjustment factor to generate a thermal inertia compensation coefficient sequence containing time delay correction.

[0150] The binary control flow and standardized thermal inertia dynamic adjustment factor in the control signal are obtained as input conditions. The binary control flow is parsed at the bit level, and a logic state matrix is ​​established to accurately identify transient operating conditions that meet the characteristics of a sudden drop in temperature forecast and slow local temperature drop.

[0151] For the identified transient operating conditions, a preset thermal inertia delay response algorithm is invoked, with the standardized thermal inertia dynamic adjustment factor as the core input parameter, and the heat transfer model of the greenhouse envelope and the specific heat capacity data of the heat storage medium are loaded to construct a heat transfer delay discrimination function.

[0152] In the discriminant function, the phase lag compensation is defined based on the difference in temperature change rates, and the time delay correction is calculated using the following formula: in, This is the time delay correction amount. The thermal inertia retardation coefficient, This refers to the predicted temperature change range. This represents the amplitude of locally measured temperature changes.

[0153] The calculated time delay correction is nonlinearly superimposed with the standardized thermal inertia dynamic adjustment factor, and then smoothed using a hyperbolic tangent function to eliminate the influence of abrupt boundary on load response and generate a continuous compensation curve.

[0154] Discretization sampling and time-series indexing are performed on the compensation curve. The thermal inertia compensation values ​​at different time steps are assembled into a thermal inertia compensation coefficient sequence in sequence. Physical consistency constraints are applied during the sequence generation process to ensure that the coefficient changes do not exceed the preset physical limits.

[0155] Through the above chain processing method, the control signal of the previous step is transformed into a thermal inertia compensation coefficient sequence that can be directly applied in the load prediction model, thereby realizing phase lag feedforward correction for situations where the temperature drops sharply and the local cooling is slow.

[0156] For example, in a greenhouse equipped with a double-layered inflatable membrane covering, an internal water-containing wall heat storage medium, and crops in flowering stage, the transient operating mode corresponding to the input binary control flow "101" is analyzed and identified as a predicted temperature drop of 3.2℃, with a local measured drop of only 0.8℃, and a standardized thermal inertia dynamic adjustment factor of 0.65. The thermal inertia delay response algorithm is invoked, the thermal inertia delay coefficient k is set to 0.15, and the time delay correction is calculated using the formula, resulting in... The time is 0.1125 hours. The correction value is superimposed with the adjustment factor of 0.65 and smoothed by the hyperbolic tangent function to obtain a compensation value of 0.685. A thermal inertia compensation coefficient sequence of length 8 is generated by sampling, ranging from 0.68 to 0.70. This sequence is superimposed on the original load prediction value in the subsequent prediction model, which significantly reduces the response speed of the model to changes in sensible heat load under this condition and avoids premature cooling start caused by misprediction.

[0157] S6.2: Obtain the crop transpiration correction coefficient in the control signal and the humidity deviation component in the fused feature vector. Perform abnormal pattern matching processing on the humidity deviation component, which is characterized by continuous warming and sudden increase in local humidity. Activate the crop transpiration latent heat gain model based on the successfully matched abnormal pattern. Perform nonlinear weighted correction on the initial sensible heat load estimate using the crop transpiration correction coefficient to generate a dynamic load correction vector containing a latent heat load enhancement term.

[0158] The crop transpiration correction coefficient contained in the control signal and the humidity deviation component in the fused feature vector are obtained. The abnormal mode detection unit is called to perform local humidity state identification on the humidity deviation component and generate a humidity change time series matrix.

[0159] The humidity change time series matrix is ​​filtered for a continuously warm background condition to remove time periods with low temperature or insufficient fluctuation frequency, so as to ensure that the abnormal pattern matching operates under stable warm conditions.

[0160] Based on the filtered humidity change time series matrix, a humidity anomaly pattern feature vector is constructed, and a similarity matching operation is performed with a pre-set anomaly pattern database to obtain a matching score vector.

[0161] Set a matching degree threshold judgment condition. When there are entries in the matching degree score vector that exceed the preset threshold, the activation signal of the crop transpiration latent heat gain model is triggered.

[0162] In the crop transpiration latent heat gain model, the initial sensible heat load estimate is nonlinearly weighted and corrected using a transpiration enhancement factor. The correction formula is as follows: in, This is the initial estimated sensible heat load. This is the crop transpiration correction factor. This is the constant for the latent heat gain term.

[0163] The corrected heat load value is encapsulated into a dynamic load correction vector by time index and output to the subsequent multi-source heterogeneous data alignment and time-series synchronization processing module. The latent heat term in the load forecast is enhanced and compensated through nonlinear weighted correction.

[0164] By using humidity anomaly pattern matching and latent heat gain model weighted correction processing, the control signal from the previous step is transformed into a dynamic load correction vector containing latent heat load enhancement terms, thereby improving the physical consistency and robustness of load trend prediction under conditions of continuous warming and sudden increase in local humidity.

[0165] For example, in a cluster of greenhouses for facility agriculture, the crop transpiration correction coefficient of 0.35 is derived from the measured value of the crop canopy transpiration rate, and the humidity deviation component is 12.5. The abnormal pattern database defines a warmer-than-normal and sudden increase in humidity condition as a humidity deviation component exceeding 10 and lasting for at least 20 minutes. The pattern matching score reaches 0.82, which is greater than the matching threshold of 0.8, triggering the latent heat gain model. The latent heat gain term constant is set to 150, and the corrected load value is calculated using the formula. =Q + 0.35 × 150, with a correction value of 552.5 under the condition of Q = 500. In actual operation, after this correction vector is input into the multi-source data fusion module, the estimation of sensible heat load in the predictive control response of the heat pump unit is significantly improved, avoiding the problem of underestimation of latent heat caused by high humidity, and improving the consistency and stability of system prediction and control.

[0166] S6.3: Obtain the thermal inertia compensation coefficient sequence and the dynamic load correction vector, perform multi-source heterogeneous data alignment and time-series synchronization processing on both, construct a feedforward offset correction function based on the synchronized data, and use the feedforward offset correction function to superimpose and fuse the time delay effect of the thermal inertia compensation coefficient sequence and the latent heat enhancement effect of the dynamic load correction vector in the time domain to generate a load trend offset correction amount with physical consistency constraints.

[0167] Obtain the synchronous data index of the thermal inertia compensation coefficient sequence and the dynamic load correction vector, and perform precise matching on the time axis to ensure that the timestamp correspondence between the two sequences is strictly consistent with the sampling interval.

[0168] The two synchronized sequences are processed to unify the multi-source heterogeneous data format, matching the numerical properties of the thermal inertia compensation coefficient sequence with the unit system of the dynamic load correction vector to the same physical dimension, and eliminating the scale offset introduced by the difference in data sources.

[0169] A feedforward offset correction function is constructed, and the time delay effect of the thermal inertia compensation coefficient and the latent heat enhancement effect of the dynamic load correction vector are fused in the time domain using a convolution integral form. Physical consistency constraints are introduced during the fusion process, and amplitude limiting and phase smoothing are applied to the superposition result to ensure that the rate of change of the offset correction in the time domain conforms to greenhouse thermal inertia and crop transpiration characteristics.

[0170] The offset correction amount, which is subject to physical consistency constraints, is mapped back to the output interface of the original load forecasting model to form an intermediate amount that can be directly used for subsequent load curve correction.

[0171] By using multi-source heterogeneous data alignment, physical dimension matching, convolutional integral fusion, and physical constraint optimization, the thermal inertia compensation coefficient sequence and dynamic load correction vector from the previous step are transformed into load trend offset correction quantities with physical consistency constraints, thereby achieving a comprehensive feedforward correction effect for meteorological errors, thermal inertia delay, and latent heat enhancement.

[0172] S6.4: Obtain the original heat load prediction curve output by the basic load prediction model and the load trend offset correction amount. Perform point-by-point linear superposition and boundary truncation processing on the original heat load prediction curve. Set physical feasible region constraints based on the heat transfer limit of the greenhouse enclosure structure and the upper limit of the heat storage medium capacity. Use the physical feasible region constraints to optimize the amplitude clamping and slope smoothing of the superimposed load curve to generate the heat load prediction value under the physical feasible region constraints.

[0173] The original heat load forecast curve output from the basic load forecast model and the load trend offset correction amount with physical consistency constraints are obtained. A point-by-point linear superposition process is performed on the original heat load forecast curve, and the correction amount is superimposed onto the original curve at each time step to form a preliminary corrected curve. Boundary truncation is performed on the preliminary corrected curve. Physical feasible region constraints are set based on the heat transfer limit of the greenhouse envelope and the upper limit of the heat storage medium capacity. Upper and lower limit truncation rules are constructed to ensure that the curve amplitude does not exceed the physical safety range. Based on the truncated curve, amplitude clamping operation is performed to ensure that the curve meets the amplitude constraints throughout the time domain. Simultaneously, a slope smoothing optimization algorithm is used to smooth the rate of change of adjacent points of the curve, eliminating the risk of control instability caused by abrupt changes. A smoothing operator based on second-order difference is used to reduce curve fluctuations while retaining the main characteristics of trend changes, forming a heat load forecast value that meets the physical feasible region constraints and has controllable stability.

[0174] By employing the aforementioned feedforward superposition, delay correction, truncation clamping, and slope optimization methods, the load trend offset correction from the previous step is transformed into a predictive output that satisfies the safety constraints of the greenhouse thermal inertia boundary and meteorological uncertainty, thereby improving the robustness of the heat pump unit control model under extreme weather conditions.

[0175] For example, in a scenario involving a cluster of multi-span glass greenhouses, the original heat load prediction curve output by the basic load prediction model reaches a maximum of 32.8 kW on a cold wave day in winter, and the trend offset correction with physical consistency constraints reaches +4.5 kW during the corresponding period. The correction is then added point-by-point to each sampling time point of the original curve, resulting in a preliminary corrected curve with a maximum value of 37.3 kW. The heat transfer limit of the greenhouse envelope structure, calculated by thermal engineering, is 35.0 kW, and the continuous power corresponding to the upper limit of the hot water storage wall capacity is 34.5 kW. Physical feasible domain constraints are set at an upper limit of 34.5 kW and a lower limit of 10.0 kW. Boundary truncation is applied to the preliminary corrected curve, directly cutting off the portion exceeding 34.5 kW to 34.5 kW. During the amplitude clamping calculation, the following amplitude constraint formula is applied: in, This is the original predicted load value. This is the trend offset correction amount. This represents the upper bound of the physically feasible region. Slope smoothing optimization is performed on the truncated curve using a second-order difference smoothing formula: in, The value is a second-order difference. After smoothing, the maximum value of the curve stabilizes at 34.5kW, and the rate of change between adjacent sampling points is less than 0.8kW / min. The final output heat load prediction value shows that the heat pump unit operates smoothly, energy utilization is greatly improved, and there is no overload or underload phenomenon in the execution environment, realizing the stable operation of the greenhouse heat pump control system under extreme weather conditions.

[0176] S6.5: Obtain the predicted heat load value under the physical feasible domain constraint and its corresponding uncertainty boundary marker, perform confidence encapsulation processing on the uncertainty boundary marker, construct a structured load prediction data packet with fault tolerance identifier based on the encapsulated data, and output the structured load prediction data packet to the downstream heat pump unit control module to complete the full-link closed-loop correction from meteorological error perception to accurate load prediction.

[0177] Obtain the predicted heat load value under the physical feasible domain constraint generated by step S6.4 and its corresponding uncertainty boundary marker data. Perform confidence encapsulation operation based on multi-dimensional parameters on the uncertainty boundary data. Encode the quantitative indicators of the boundary marker information and the amplitude characteristics of the predicted value in matrix form to form a unified data structure that can be called for subsequent decision-making.

[0178] Multi-scale statistical feature extraction is performed on the boundary-marked data of the encapsulated computation input, including mean shift, variance inflation coefficient, and kurtosis parameter. A confidence quantification function is constructed using a mapping formula, binding the predicted values ​​at different locations within the physical feasible region with their corresponding confidence weights in structured rows and columns. The formula is as follows: in, This represents the confidence coefficient after encapsulation. For the predicted value amplitude, For the boundary center value, The standard deviation of the boundary distribution is given.

[0179] The confidence coefficients and predicted values ​​are combined to form a two-dimensional pairing matrix, and cross-time domain index mapping is performed so that the position of the predicted data on the time axis can correspond one-to-one with the position of the confidence boundary.

[0180] The predicted data, which has been encapsulated with confidence levels, is subjected to compliance verification with the physical feasible domain constraints. Numerical values ​​that exceed the constraint range are labeled as fault-tolerant, thereby constructing a structured load prediction data package with fault-tolerant identification.

[0181] The structured data packets are formatted and protocolized, and each field is mapped to the interface requirements of the downstream heat pump unit control module to ensure that the output data is correctly parsed at both the semantic and physical layers.

[0182] By using confidence encapsulation and boundary correlation processing, the prediction results of step S6.4 are transformed into structured load prediction data with identifiable fault-tolerant states and verifiable physical constraints, thereby enabling the secure transmission and execution of load prediction in the error perception and thermal inertia compensation closed loop.

[0183] For example, in a 300㎡ glass greenhouse, the predicted heat load under the physical feasible region constraint output by step S6.4 is 8.5kW, the center of uncertainty boundary value is 8.0kW, and the standard deviation is 0.3kW. When performing the confidence encapsulation calculation, take... =8.5, =8.0, =0.3, calculated using the corresponding formula: The confidence coefficient was approximately 4.17. A fault-tolerance flag was assigned to values ​​exceeding the physical feasible region, ensuring that the data packets received by the control module explicitly included the predicted value, confidence coefficient, and fault-tolerance status. In this scenario, the heat pump unit was able to delay the heating process based on the fault-tolerance flag, effectively avoiding the risk of short-term overload caused by prediction errors. During trial operation, the system response delay was reduced to less than 35 seconds, the prediction curve remained smooth, and no safety boundary exceedances were triggered.

[0184] Step S7: Based on the predicted heat load value, and combined with the error range constrained by the forecast confidence label and the thermal inertia parameter set, generate a preheating or precooling start / stop control command for the heat pump unit. Specifically, this includes: It should be noted that the error range refers to the load prediction uncertainty boundary determined by the error magnitude and direction information in the forecast confidence label and the load attenuation coefficient in the thermal inertia parameter set; this boundary is specifically quantified in S7.1 as a heat load safe operation threshold vector, which is used to constrain the generation of heat pump unit start-up and shutdown control commands to ensure that the control commands do not exceed the greenhouse physical feasible domain.

[0185] S7.1: Obtain the predicted heat load value under the physical feasible region constraint and its corresponding uncertainty boundary marker data, and perform confidence interval analysis and extreme value extraction processing on the uncertainty boundary marker data to generate a heat load safe operation threshold vector characterizing the allowable range of load fluctuation.

[0186] The predicted heat load under the constraints of the physical feasible region is obtained as the input basis for this step, and the uncertainty boundary marker data generated by the previous step is obtained as additional parameters.

[0187] The uncertainty boundary marker data is subjected to confidence interval parsing processing, which splits the upper and lower limit information in the boundary markers according to the time index and transforms them into quantifiable interval endpoint vectors.

[0188] The difference sequence between the predicted value and the interval boundary is calculated using the interval endpoint vector, and the difference sequence is subjected to symbol classification to distinguish between the state categories of exceeding the upper limit, exceeding the lower limit, and within the allowable range.

[0189] The extreme value extraction module is invoked to extract the maximum allowable positive offset and the maximum allowable negative offset corresponding to each time step in the difference sequence, and form a double-ended limiting matrix composed of the extreme values ​​at both ends.

[0190] Normalization is used to scale the amplitude of the double-ended limiting matrix according to the heat transfer limit and heat storage capacity constraints of the physical feasible domain, so that the limiting data matches the power regulation capability of the heat pump unit.

[0191] Through the above processing method, the predicted value and uncertainty boundary marker of the previous step are transformed into a heat load safe operation threshold vector that characterizes the allowable range of load fluctuation, thereby supporting the stability and fault tolerance of unit regulation.

[0192] For example, during the operation of a greenhouse on a certain day, the predicted heat load under the physical feasible region constraint is 45kW, 48kW, and 50kW at different time steps, with corresponding upper and lower limits of uncertainty boundary markings of ±3kW, ±2.5kW, and ±4kW, respectively. After confidence interval resolution, the interval endpoint vector is obtained as [(42kW,48kW),(45.5kW,50.5kW),(46kW,54kW)]. The difference sequences are (+0kW), (-0.5kW), and (-4kW), respectively, and the sign classification results are within the allowable range, above the lower limit, and within the allowable range, respectively. The extreme value extraction module extracts the maximum allowable positive offset and negative offset matrix [(+3kW,-3kW),(+2.5kW,-2.5kW),(+4kW,-4kW)]. Under the conditions of a heat storage capacity of 60kW and a heat transfer limit of 5kW, normalization processing was performed by scaling each value in the limiting matrix by 5kW to obtain the corresponding heat load safe operation threshold vector [(0.6,-0.6),(0.5,-0.5),(0.8,-0.8)]. This threshold vector is used in the control strategy to limit the adjustment range of the unit's output power. Actual measurements showed that under the above conditions, the frequency of start-ups and shutdowns and operational fluctuations were significantly reduced, and the stability was significantly improved.

[0193] S7.2: Based on the heat load safe operation threshold vector and the current measured ambient temperature data inside the greenhouse, perform deviation calculation and trend prediction processing, and use the proportional-integral-derivative adjustment algorithm to perform dynamic compensation calculation on the prediction deviation to generate an initial unit power demand signal containing feedforward correction.

[0194] S7.3: Obtain the initial unit power demand signal and the current operating status parameters of the heat pump unit, perform start-stop frequency statistics and minimum operating time verification on the operating status parameters, and construct an anti-oscillation logic judgment matrix in combination with the heat load safe operation threshold vector to generate a smoothed and optimized set of unit action candidate instructions.

[0195] The initial unit power demand signal, including feedforward correction, and the current operating status parameters of the heat pump unit are obtained. Time index association processing is performed on these operating status parameters, precisely matching the time-series label of the power demand signal with the start-stop state history, cumulative operating time, and transient power output records in the operating status parameters. Start-stop frequency statistics are performed on the matched start-stop state history, using a sliding window method to extract the number of start-stops per unit time, and comparing it differentially with a preset safe start-stop frequency threshold to form a start-stop frequency deviation index. Based on the start-stop frequency deviation index, a minimum operating time check is performed on the cumulative operating time, constructing a logical judgment branch. When the expected operating time of a candidate start-stop action is less than the minimum operating time threshold, it is marked as not executed. Combining the heat load safe operating threshold vector, the start-stop frequency deviation index and the minimum operating time check result are matrix-encoded to form an anti-oscillation logic judgment matrix. Threshold upper and lower limit constraint rules are embedded in this matrix to suppress high-frequency start-stops caused by prediction fluctuations. Under the action of the anti-oscillation logic judgment matrix, the initial unit power demand signal is smoothed and its amplitude gradually varied, eliminating sudden power demands that do not meet stability conditions, and generating a set of candidate unit action instructions that has been smoothed and optimized and meets the requirements of operational continuity. Through the above processing method, the power demand signal of the previous step is transformed into a stable instruction set that can avoid operational oscillations, achieving high fault tolerance and high stability of heat pump unit operation in weather forecast error compensation scenarios.

[0196] For example, the heat pump unit configured in the greenhouse management system has a rated power of 30kW, a safe start-stop frequency threshold of no more than 3 times per hour, and a minimum operating time threshold of 15 minutes. During a certain detection cycle, the initial unit power demand signal exhibits several drastic changes. The collected operating status parameters include short cycles of 4 start-stop operations performed within the last hour, with single operating times of 8 minutes and 12 minutes respectively. After associating the power demand signal with the operating status parameters, the system extracts the number of start-stop operations per hour as 4 using a sliding window. Comparing this to the threshold of 3, the deviation is 1, and this state is marked as prohibited during matrix encoding. Candidate actions with a cumulative operating time of less than 15 minutes are directly eliminated based on logical judgment branches. The upper limit power recorded in the heat load safe operating threshold vector is 28kW, and the lower limit power is 12kW. After embedding this constraint into the matrix, smoothing filtering adjusts the sudden increase in power demand signal from 28.5kW to 28kW and the sudden decrease from 11.5kW to 12kW. The final output set of candidate unit action instructions includes two executable power adjustment instructions, which correspond to a stable increase in power to 26kW and a stable decrease in power to 14kW, respectively. The verification results show that after the execution of this instruction set, the unit did not experience high-frequency start-up and shutdown, and the fluctuation of the power output curve was significantly reduced.

[0197] S7.4: Based on the upper and lower limits of the candidate unit action instruction set and the upper and lower limits of the heat load safe operation threshold vector, perform multi-objective optimization screening and priority sorting processing, and replace aggressive control instructions that exceed the physical feasible domain with boundary saturation values ​​to generate a standardized heat pump unit pre-adjustment control sequence with fault tolerance.

[0198] Based on the smoothed and optimized candidate instruction set of the unit actions and the upper and lower limit constraints in the thermal load safe operation threshold vector, the multi-objective optimization solution module is called to perform quantitative screening calculations on each candidate instruction set. The performance indicators of the unit power demand meeting load tracking, the start-stop frequency constraint indicators, and the energy efficiency optimization indicators are weighted and combined in a unified evaluation function to form a comparable comprehensive scoring matrix.

[0199] Boundary constraint elimination processing is performed on the comprehensive scoring matrix. Candidate instructions in the scoring that involve power exceeding the upper limit threshold or falling below the lower limit threshold are marked as out-of-bounds states. The out-of-bounds power values ​​are replaced with the corresponding threshold constants through the boundary saturation value replacement mechanism to ensure that the control sequence does not violate the physical feasible domain logic.

[0200] Priority sorting is performed on the boundary-corrected scoring matrix. The timing sensitivity of the strategy execution is weighted and corrected using a priority calculation formula based on the thermal response time constant and the current prediction deviation trend. The sorting result is used to determine the execution order and time interval of each instruction in the pre-adjustment control sequence.

[0201] A standardized mapping function is used to encode the top-ranked candidate instructions that have undergone boundary saturation correction into a unified control sequence structure. The structure includes the unit power setpoint, execution timestamp, and control mode identifier fields to ensure that the downstream controller can directly parse and implement the action instructions.

[0202] By performing multi-objective optimization screening, boundary saturation value replacement, and priority sorting, the results of the previous step are transformed into a standardized heat pump unit pre-adjustment control sequence with fault tolerance, achieving a control output that balances physical feasibility and energy efficiency optimization while also being fault-tolerant.

[0203] For example, in a greenhouse cluster operation scenario, the input set of candidate unit action commands includes power demand signals of 5.5kW, 6.8kW, and 8.2kW, and a heat load safe operation threshold vector with an upper limit of 7.0kW and a lower limit of 5.0kW, a thermal response time constant of 2 hours, and a prediction deviation trend of increasing positive deviation. The multi-objective optimization evaluation function sets weight ratios of 0.5 for load tracking performance, 0.3 for start-stop frequency constraints, and 0.2 for energy efficiency optimization, forming a comprehensive scoring matrix. For the 8.2kW command, exceeding the upper threshold, a boundary saturation value replacement mechanism is used to correct the power to 7.0kW. During the priority ranking process, the thermal response time constant of 2 hours and the increasing prediction deviation trend are considered, using the priority formula: in This is the overall score. To predict the magnitude of the deviation change, The thermal response time constant is used. Calculations show that the 6.8kW command has the highest priority in the sequence, followed by the modified 7.0kW command. In the final generated standardized pre-regulation control sequence, the first command is 6.8kW power, with an execution timestamp corresponding to the current cycle and a mode identifier of heating; the second command is 7.0kW power, delayed by one execution cycle, and also identified as heating. In actual execution verification, this sequence maintained the unit operation within safe thresholds, demonstrated stable load tracking, and significantly improved energy efficiency.

[0204] S7.5: Obtain the standardized heat pump unit pre-adjustment control sequence, perform communication protocol encapsulation and timing synchronization marking processing on the standardized heat pump unit pre-adjustment control sequence, and convert it into a binary start-stop control command stream that can directly drive the heat pump controller hardware, so as to complete the full-link closed-loop control output from load forecasting to unit execution.

[0205] Step S8: Obtain the execution result of the heat pump unit preheating or precooling start / stop control command, and update the statistical distribution characteristics of the forecast error vector sequence based on the deviation between the execution result of the heat pump unit preheating or precooling start / stop control command and subsequent actual meteorological data. Specifically, this includes: S8.1: Obtain the actual temperature response curve inside the greenhouse after the execution of the preheating or precooling start-stop control command of the heat pump unit, the public weather forecast service data for the subsequent period, and the actual local micro-meteorological data collected by the low-cost micro-weather station array. Perform strict time axis alignment and abnormal noise removal processing on the actual temperature response curve inside the greenhouse, the public weather forecast service data for the subsequent period, and the actual local micro-meteorological data to generate a high-precision spatiotemporal synchronous measured benchmark dataset for closed-loop verification.

[0206] S8.2: Based on the high-precision spatiotemporal synchronous measured benchmark dataset, extract the actual local temperature and humidity signals and actual wind speed signals at the corresponding times, and perform element-wise difference operations on them with the original public meteorological forecast service data contained in the high-precision spatiotemporal synchronous measured benchmark dataset to reconstruct a real-time posterior meteorological forecast error vector sequence that reflects the latest forecast deviation characteristics.

[0207] S8.3: Obtain the real-time posterior meteorological forecast error vector sequence and the historically stored forecast error vector sequence, and perform sliding window statistical moment calculation and probability density function fitting on both to quantitatively analyze the mean drift, variance expansion coefficient and skewness distribution evolution trend of the temperature deviation component, humidity deviation component and wind speed deviation component, thereby generating an error distribution evolution feature matrix that characterizes the dynamic changes of the statistical characteristics of forecast errors.

[0208] S8.4: Based on the error distribution evolution feature matrix, the Bayesian online parameter update algorithm is called to iteratively correct the prior distribution parameters in the preset meteorological forecast error dynamic perception module, and the mean drift, variance expansion coefficient and skew distribution evolution trend are integrated into the prior distribution parameters to generate a corrected meteorological forecast error statistical distribution model with adaptive update capability.

[0209] S8.5: Replace the original statistical distribution feature definition of the forecast error vector sequence with the modified weather forecast error statistical distribution model, and re-inject the updated statistical distribution features into the error identification input of the weather forecast error dynamic perception and thermal inertia compensation coupled model to complete the full-link closed-loop iteration from execution result feedback to model parameter optimization, and ensure continuous improvement of the model's robustness under varying operating conditions.

[0210] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0211] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0212] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered 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 pre-adjusting the load of greenhouse heat pump units based on outdoor weather forecasts, specifically including: S1: Acquire real-time changes in local temperature and humidity, wind speed and public weather forecast data, and compare the real-time changes in local temperature and humidity and wind speed with the public weather forecast data to generate a forecast error vector sequence; S2: Perform confidence quantification based on the forecast error vector sequence to generate forecast confidence labels; S3: For greenhouse buildings with different structural types, different covering materials and different internal heat storage configurations, a set of thermal inertia parameters is constructed using the thermal response time constant, temperature hysteresis phase angle and load attenuation coefficient obtained from offline calibration; S4: Obtain conventional weather forecast elements, take the time series features of the conventional weather forecast elements as the main input, and take the forecast confidence label and the corresponding thermal inertia parameter set index as auxiliary input, perform feature fusion processing, and generate a fused feature vector; S5: Based on the abrupt change intensity of the prediction error vector in the fused feature vector, dynamically adjust the thermal inertia compensation weight to generate a control signal; S6: Use the control signal to perform advance deviation correction processing on the load change trend. When a sudden drop in temperature forecast is detected and the local measured temperature drop is slow, activate the thermal inertia delay response. Or when the forecast is continuously warmer and the local humidity rises sharply, call the crop transpiration correction coefficient to generate the heat load prediction value. S7: Based on the predicted heat load value, and combined with the error range constrained by the forecast confidence label and the thermal inertia parameter set, generate a heat pump unit preheating or precooling start / stop control command.

2. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecasting according to claim 1, characterized in that, Step S7 is followed by step S8, which specifically includes: S8: Obtain the execution result of the preheating or precooling start-stop control command of the heat pump unit, and update the statistical distribution characteristics of the forecast error vector sequence based on the deviation between the execution result of the preheating or precooling start-stop control command of the heat pump unit and the subsequent actual meteorological data.

3. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecasting according to claim 1, characterized in that, Step S4 specifically includes: Obtain conventional weather forecast elements and their time series characteristic data, perform sliding window statistics calculation and frequency domain transformation processing on the conventional weather forecast elements and their time series characteristic data, and generate the main channel time series characteristic matrix; Based on the time index of the main channel time series feature matrix, a preset mapping retrieval mechanism is invoked to extract the corresponding forecast confidence label and thermal inertia parameter set index. One-hot encoding mapping and numerical normalization processing are performed on the forecast confidence label and thermal inertia parameter set index to generate the auxiliary channel state vector. The attention mechanism weight allocation algorithm is used to perform cross-modal feature interaction operation on the main channel time series feature matrix and the auxiliary channel state vector. The feature weights of each time step of the main channel are adjusted according to the error fluctuation intensity in the forecast confidence label to generate a weighted meteorological feature tensor. Based on the weighted meteorological feature tensor and the thermophysical information in the auxiliary channel state vector, feature splicing and nonlinear activation function mapping are performed to bind the meteorological error propagation path and the greenhouse thermal inertial response delay characteristics in the feature space, generating an intermediate feature vector. The intermediate feature vector is subjected to dimensionality reduction projection and orthogonalization cleaning to remove redundant feature components and unify feature dimensions, thereby generating a fused feature vector.

4. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecast as described in claim 1, characterized in that, Step S5 specifically includes: The prediction error vector component in the fused feature vector is obtained, and the prediction error vector component is subjected to sliding window second-order difference operation and gradient magnitude calculation to generate a scalar sequence of error mutation intensity. Based on the error mutation intensity scalar sequence, an adaptive threshold segmentation process is performed by calling a pre-set nonlinear mapping function library to map the continuously changing error mutation intensity scalar sequence into a discretized meteorological uncertainty level index, thereby generating a classification status identifier. The thermal response time constant and temperature hysteresis phase angle corresponding to the classification state identifier are extracted from the thermal inertia parameter set. A dynamic weight allocation model is constructed using the hyperbolic tangent activation function. Nonlinear weighted fusion operation is performed on the thermal response time constant and temperature hysteresis phase angle to generate initial thermal inertia compensation weight coefficients. Using the initial thermal inertia compensation weight coefficient, combined with the crop transpiration correction coefficient in the fused feature vector, the proportional-integral-derivative adjustment algorithm is executed to perform phase lag compensation correction, and the corrected weight coefficient is subjected to normalization constraint processing to generate a standardized thermal inertia dynamic adjustment factor. Based on the standardized thermal inertia dynamic adjustment factor and the meteorological uncertainty level index, logic gating coding and pulse width modulation signal synthesis processing are performed to convert the numerical adjustment factor into a binary control flow that can drive the switching states of neurons inside the load prediction model, thereby generating the control signal.

5. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecast as described in claim 1, characterized in that, Each component in the forecast error vector is normalized using its historical statistical maximum value and physical possible range. The Euclidean distance of the normalized vector is calculated as the error magnitude. The arctangent function is used to calculate the error direction angle. The magnitude and direction angle are concatenated to form the forecast confidence label.

6. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecast as described in claim 1, characterized in that, In the process of constructing the thermal inertia parameter set, the thermal response time constant is obtained through dynamic thermal balance simulation, the temperature hysteresis phase angle is obtained through Fourier transform, and the load attenuation coefficient is obtained through envelope fitting. The three parameters are associated with and normalized with the greenhouse structure type, covering material, and crop stage characteristics, and a unique index is established.

7. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecast as described in claim 1, characterized in that, By utilizing the attention mechanism and the confidence label of the auxiliary channel, an error fluctuation intensity parameter is introduced into the weighted allocation calculation of feature fusion from the main channel to the auxiliary channel. The feature contribution of multi-time step input is dynamically adjusted through softmax temperature regulation to suppress the meteorological input weight during extreme fluctuations.

8. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecast as described in claim 1, characterized in that, The control signal is synthesized by logic gating encoding and pulse width modulation signal, which maps the meteorological uncertainty level and adjustment factor to different PWM duty cycles, and maps them to binary signal streams that can drive different switching states of model neurons.

9. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecast as described in claim 1, characterized in that, When generating the start-stop control command for the heat pump unit, an anti-oscillation objective function is constructed based on comprehensive indicators such as the heat load safe operation threshold, the current state of the unit, and the minimum operating time. After candidate commands undergo multi-objective optimization, boundary truncation, and priority sorting, a control sequence is output.

10. The method for pre-adjusting the load of a greenhouse heat pump unit based on outdoor weather forecast as described in claim 2, characterized in that, Based on the deviation between the execution result of the control command and the actual meteorological data, the posterior error is calculated after high-precision alignment and anomaly removal, and the prior distribution of meteorological error is updated using a Bayesian online correction algorithm.

Images