Warehousing temperature and humidity intelligent regulation and control method and system based on deep learning

Abstract

The invention discloses a storage temperature and humidity intelligent regulation and control method and system based on deep learning, and relates to the field of artificial intelligence, and the method comprises the steps: firstly inputting current temperature and humidity time sequence data and an associated equipment operation log into a target temperature and humidity regulation and control model comprising a feature extraction module and a strategy generation module; the temperature and humidity time sequence data and equipment operation logs are coded and integrated into a fusion environment feature set through a feature extraction module; after a regulation and control strategy feature set is obtained, a strategy generation module executes gating circulation processing on fusion features and strategy features, and a gating circulation data set representing the adaptation relation is output; and outputting an adaptation probability based on the data set, and screening a target regulation and control strategy. According to the invention, accurate matching of the temperature and humidity dynamic state and the regulation strategy is realized through deep learning, the timeliness and energy efficiency ratio of storage temperature and humidity regulation are improved, and the stability of a goods storage environment is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a method and system for intelligently controlling storage temperature and humidity based on deep learning. Background Art

[0002] Current warehouse temperature and humidity control methods rely heavily on manual experience or fixed rules, resulting in delayed response, low energy efficiency, and poor adaptability to complex scenarios. As e-commerce warehouses expand and the sensitivity of goods stored increases, traditional control methods are unable to meet the demand for dynamic and precise temperature and humidity control. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for intelligent control of storage temperature and humidity based on deep learning.

[0004] In a first aspect, an embodiment of the present invention provides a method for intelligently controlling storage temperature and humidity based on deep learning, comprising:

[0005] Inputting the current temperature and humidity time series data and the current equipment operation log associated with the current temperature and humidity time series data into a target temperature and humidity control model; the target temperature and humidity control model includes a feature extraction module and a strategy generation module;

[0006] Performing feature encoding on the current temperature and humidity time series data and the current device operation log by the feature extraction module in the target temperature and humidity control model to obtain a temperature and humidity time series feature set corresponding to the current temperature and humidity time series data, and a current device operation log feature corresponding to the current device operation log; performing a feature integration operation on the temperature and humidity time series feature set and the current device operation log feature to obtain a fused environment feature set;

[0007] Acquire control strategy feature sets corresponding to multiple control strategies, perform gated cycle processing on the fused environment feature set and the control strategy feature set through the strategy generation module, and obtain a gated cycle data set corresponding to the control strategy feature set; each gated cycle value in the gated cycle data set is used to characterize the adaptation relationship between the current temperature and humidity time series data and a control strategy;

[0008] The adaptation probability values ​​corresponding to the multiple control strategies are output according to the gated cycle data set, and the target control strategy among the multiple control strategies is set for the current temperature and humidity time series data according to the adaptation probability values.

[0009] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is configured to execute the method described in the first aspect.

[0010] Compared with the existing technology, the beneficial effects provided by the present invention include: using a deep learning-based intelligent storage temperature and humidity control method and system disclosed by the present invention, by inputting the current temperature and humidity time series data and the associated equipment operation log into a target temperature and humidity control model containing a feature extraction module and a strategy generation module; encoding the temperature and humidity time series data and the equipment operation log through the feature extraction module and integrating them into a fusion environment feature set; after obtaining the control strategy feature set, the strategy generation module performs gated loop processing on the fusion features and the strategy features, and outputs a gated loop data set that characterizes the adaptation relationship; based on the data set, the adaptation probability is output and the target control strategy is screened. The present invention achieves precise matching of temperature and humidity dynamics with control strategies through deep learning, improves the timeliness and energy efficiency of warehouse temperature and humidity control, and ensures the stability of the cargo storage environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly describes the drawings required for use in the embodiments. It should be understood that the following drawings illustrate only certain embodiments of the present invention and should not be construed as limiting the scope of the present invention. Those skilled in the art can, without inventive effort, derive other relevant drawings from these drawings.

[0012] Figure 1 A schematic diagram of the steps of the intelligent control method of storage temperature and humidity based on deep learning provided by an embodiment of the present invention;

[0013] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the described embodiments are only a portion of the embodiments of the present invention, not all of them. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

[0015] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0016] In order to solve the technical problems in the above background technology, Figure 1 This is a flow chart of the method for intelligently controlling storage temperature and humidity based on deep learning provided in an embodiment of the present disclosure. The method for intelligently controlling storage temperature and humidity based on deep learning is introduced in detail below.

[0017] Step S201: inputting current temperature and humidity time series data and current equipment operation log associated with the current temperature and humidity time series data into a target temperature and humidity control model; the target temperature and humidity control model includes a feature extraction module and a strategy generation module;

[0018] Step S202: Feature encoding is performed on the current temperature and humidity time series data and the current device operation log by the feature extraction module in the target temperature and humidity control model to obtain a temperature and humidity time series feature set corresponding to the current temperature and humidity time series data, and a current device operation log feature corresponding to the current device operation log. Feature integration is performed on the temperature and humidity time series feature set and the current device operation log feature to obtain a fused environment feature set.

[0019] Step S203: obtaining control strategy feature sets corresponding to multiple control strategies, performing gated cycle processing on the fused environment feature set and the control strategy feature set by the strategy generation module to obtain a gated cycle data set corresponding to the control strategy feature set; each gated cycle value in the gated cycle data set is used to characterize the adaptation relationship between the current temperature and humidity time series data and a control strategy;

[0020] Step S204 : outputting adaptation probability values ​​corresponding to the plurality of control strategies respectively according to the gated cycle data set, and setting a target control strategy among the plurality of control strategies for the current temperature and humidity time series data according to the adaptation probability values.

[0021] In an embodiment of the present invention, for example, in an actual e-commerce warehouse temperature and humidity control scenario, the server, as the execution subject, first executes the step of "inputting the current temperature and humidity time series data and the current equipment operation log associated with the current temperature and humidity time series data into the target temperature and humidity control model". Assume that the warehouse is deployed with a high-density temperature and humidity sensor network, which collects temperature and humidity data of each area (such as the shelf area and the sorting area) every 5 minutes to form the current temperature and humidity time series data (for example, the temperature and humidity sequence of the shelf area during a certain period is [28°C, 65% RH], [29°C, 63% RH], [30°C, 62% RH]...); at the same time, the operating status of the air conditioners, dehumidifiers, fresh air systems and other equipment in the warehouse (for example, air conditioner No. 1 is running at 26°C from 9:00 to 10:00, and dehumidifier No. 2 is adjusted to a high wind speed after 10:00) is recorded by the equipment management system as the current equipment operation log. The server uses the IoT data collection interface to synchronously extract the real-time temperature and humidity time series data and the corresponding device operation logs, and inputs them into a pre-trained target temperature and humidity control model. This target temperature and humidity control model includes a built-in feature extraction module and a strategy generation module. The former is responsible for feature analysis and encoding of the input data, while the latter is used to generate control strategies based on feature matching.

[0022] Next, the server drives the feature extraction module of the target temperature and humidity control model to perform feature encoding and integration operations on the current temperature and humidity time series data and the current equipment operation log. For the current temperature and humidity time series data, the feature extraction module first divides it into multiple temperature and humidity time intervals according to the time dimension (for example, with 30 minutes as a time period, the above-mentioned sampling points every 5 minutes are aggregated into a time interval consisting of 6 sampling points), and each temperature and humidity time interval is composed of continuous temperature and humidity sampling points. For each temperature and humidity time interval, the feature extraction module performs feature mapping through a neural network structure such as a multi-layer perceptron to generate the initial features of the time period (covering basic statistical features such as the temperature mean and humidity fluctuation variance within the time period); then, the time period time series identifier is injected into each initial feature of the time period (such as the first time period and the second time period, which identifies its position order in the overall temperature and humidity time series) to obtain the enhanced features of the time period (at this time, the features not only contain temperature and humidity statistical information, but are also associated with the time series context). All time-period enhanced features are encoded using an attention mechanism or Transformer encoding to obtain the time-period encoding features corresponding to each temperature and humidity period. These time-period encoding features are then spliced ​​or weightedly fused in time sequence to generate the temperature and humidity time series feature set corresponding to the current temperature and humidity time series data (this feature set fully describes the temporal trend of storage temperature and humidity, the fluctuation pattern, and the correlation between each time period). For the current equipment operation log, the feature extraction module first deconstructs the log text into multiple log semantic units (such as "air conditioner No. 1", "9:00-10:00", "26℃ cooling", and other semantic fragments). Each log semantic unit is mapped to a semantic vector using a pre-trained word vector model (such as a BERT variant) to obtain a semantic unit vector. The log location identifier is then injected into each semantic unit vector (for example, the first semantic unit corresponds to device startup information, the second corresponds to operating parameter information) to generate a semantic enhancement vector (which integrates the semantic information with the location context within the log). All semantic enhancement vectors are sequentially encoded (e.g., using LSTM) to obtain the log semantic unit features corresponding to each log semantic unit. These features are then integrated to generate the current device operation log features corresponding to the current device operation log (this feature set covers key information such as the device's start and stop status, operating parameters, and time associations). After encoding the two feature sets mentioned above, the feature extraction module performs feature integration operations on the temperature and humidity time series feature set and the current device operation log features through feature splicing and element-by-element weighting to obtain a fused environmental feature set (this feature set also includes the time series dynamics of warehouse temperature and humidity and the real-time status of equipment operation, providing a comprehensive description of the environmental dimension for subsequent strategy matching).

[0023] The server then executes the steps of "obtaining control policy feature sets corresponding to multiple control policies" and "performing a gated loop on the fused environment feature set and control policy feature set through the policy generation module." The warehouse pre-configured multiple temperature and humidity control policies (e.g., Policy 1: "Start air conditioner 1 to cool to 24°C, and adjust the fan speed of dehumidifier 2 to medium"; Policy 2: "Turn on the fresh air system, maintain air conditioner 3 at 26°C, and turn off dehumidifier 1"). The server retrieves these control policies from the policy database and first deconstructs each policy into policy elements, including device identification, action instructions, and parameter settings (e.g., Policy 1 is deconstructed into elements such as "air conditioner 1," "start," "cool," "24°C," "dehumidifier 2," "fan speed," and "medium"). Next, the feature extraction module of the target temperature and humidity control model encodes each policy element individually: a vector representation of the device identification is generated through the embedding layer associated with the device inventory, an action vector is generated through the embedding layer of the action instruction dictionary, and a parameter vector is generated through the numerical encoding module. This yields the corresponding policy encoding feature set for each policy element. The policy encoding feature set for each policy is weighted and aggregated (e.g., a weight of 0.4 for device elements, 0.3 for action elements, and 0.3 for parameter elements) to obtain the control policy features corresponding to a single control policy. The control policy features of all control policies are then integrated according to policy numbers to form control policy feature sets corresponding to multiple control policies. The policy generation module then performs gated loop processing on the fused environment feature set and the control policy feature set. The policy generation module consists of multiple cascaded gated loop branches. For example, the first gated loop branch has its loop precursor features as the control policy feature set. The server uses these loop precursor features and the fused environment feature set as the loop input features of this branch. Within the gated loop branch, the loop precursor features (control policy feature set) are first used as the policy index vector in the gated loop operator (used to identify the core features of different control policies), and the fused environment feature set is used as the environmental sensing vector (reflecting the current warehouse temperature and humidity and equipment status) and the equipment state vector (focusing on the real-time parameters of equipment operation).The gated recurrent operator calculates the element-by-element product of the policy index vector and the environmental sensor vector to obtain the first interaction feature (which characterizes the initial association between the policy and the environment). Based on the number of temperature and humidity control directions (such as the number of cooling and dehumidifying directions, assuming there are three categories) in the fused environmental feature set, the product of the first interaction feature and the multiplicative inverse of the number of control directions is calculated to obtain the second interaction feature (which balances the influence weights of different control directions). Based on the nonlinear transformation subfunction within the gated recurrent operator (such as the ReLU activation function), the second interaction feature is converted into the first mapping feature (enhancing the nonlinear expression ability of the feature). The third interaction feature is then calculated by the element-by-element product of the first mapping feature and the device state vector to obtain the third interaction feature (which strengthens the association between the policy and the current device state). The third interaction feature is then added to the policy index vector to obtain the original gated recurrent feature (integrating multi-dimensional correlation information). The original gated recurrent features are normalized by calculating the mean and standard deviation of their neuronal activations. The difference between the original gated recurrent features and the mean is calculated as the centered feature. The standard deviation is added to a preset normalization parameter (e.g., the minimum value ε = 1e-8) and the square root is taken to obtain the normalization coefficient. Finally, the centered feature is multiplied by the multiplicative inverse of the normalization coefficient to obtain the normalized feature (eliminating dimensionality differences and improving feature stability). The normalized features are processed through a fully connected layer and an activation function (e.g., GELU) to obtain the logistic conduction feature. A residual connection operation is then performed on the normalized features and the logistic conduction features (preserving the original feature information and preventing gradient vanishing) to obtain the gated recurrent environment coupling feature. This feature is then normalized again to obtain the recurrent response feature of the first gated recurrent branch. Subsequent cascaded gated recurrent branches use the recurrent response feature of the previous branch as their own recurrent precursor feature. The above gated recurrent processing process is repeated until the last gated recurrent branch outputs the recurrent response feature, which is the gated recurrent dataset corresponding to the control strategy feature set. Each gated cycle value in the gated cycle data set quantitatively represents the adaptation relationship between the current temperature and humidity time series data and a single control strategy (the higher the value, the stronger the adaptation).

[0024] Finally, the server executes "outputting the adaptation probability values ​​corresponding to the multiple control strategies respectively according to the gated cycle data set, and setting the target control strategy among the multiple control strategies for the current temperature and humidity time series data according to the adaptation probability value". The decision output unit of the target temperature and humidity control model performs confidence normalization on the gated cycle data set based on nonlinear transformation functions such as Softmax, and converts each gated cycle value into an adaptation probability value in the range of 0-1 (such as the adaptation probability of strategy 1 is 0.85, strategy 2 is 0.62, strategy 3 is 0.78, etc.). The server presets a probability threshold (such as 0.7), filters out the control strategies whose adaptation probability values ​​exceed the threshold (such as strategy 1 and strategy 3), and determines the adaptation probability values ​​corresponding to these strategies as the target adaptation probability values. Subsequently, the server selects the control strategy with the highest target adaptation probability value (for example, 0.85 for strategy 1 is higher than 0.78 for strategy 3) as the target control strategy based on the warehouse's control priority rules (such as energy consumption priority, response speed priority, etc.), and deploys the strategy to the warehouse's equipment control system: it sends the command "start cooling and set the temperature to 24°C" to air conditioner No. 1, and sends the command "adjust the wind speed to medium" to dehumidifier No. 2, thereby realizing intelligent control of the warehouse environment corresponding to the current temperature and humidity time series data.

[0025] Throughout the entire process, the server comprehensively analyzes the temporal dynamics of warehouse temperature and humidity and the operating status of equipment through the feature extraction module of the target temperature and humidity control model. It then uses the gated loop mechanism of the policy generation module to deeply match environmental characteristics with control strategies. Ultimately, the decision output unit accurately screens and deploys target control strategies, achieving intelligent and refined control of warehouse temperature and humidity, effectively ensuring the stability of the storage environment and the energy efficiency of equipment operation. For example, in a scenario where the warehouse temperature rises sharply in the summer afternoon, the server quickly identifies the most suitable strategy, "Start air conditioning No. 1 to 24°C cooling + dehumidifier No. 2 medium wind speed," through the above process, and promptly reduces the warehouse temperature and humidity to prevent damage to goods due to excessive temperature and humidity. At the same time, precise policy matching reduces ineffective equipment operation and reduces energy costs.

[0026] In a possible implementation, the gated loop processing is performed on the fusion environment feature set and the control strategy feature set by the strategy generation module to obtain a gated loop data set corresponding to the control strategy feature set, which can be implemented through the following example.

[0027] In the strategy generation module, gated loop processing is performed on the fusion environment feature set and the control strategy feature set to obtain original gated loop features;

[0028] Normalizing the original gated cycle feature to obtain a gated cycle normalized feature;

[0029] The gated cycle normalization feature is processed through a fully connected layer and an activation function to obtain a gated cycle conduction feature, a residual connection operation is performed on the gated cycle normalization feature and the gated cycle conduction feature to obtain a gated cycle coupling feature, and the gated cycle coupling feature is standardized to obtain a gated cycle data set corresponding to the control strategy feature set.

[0030] In an embodiment of the present invention, exemplarily, when the server executes the gated loop processing of the fusion environment feature set and the control strategy feature set through the policy generation module, it first drives the gated loop operator in the policy generation module to carry out interactive calculations for the fusion environment feature set of the current warehouse (such as the temperature and humidity time series in the sorting area in the afternoon presents a fluctuation trend of 28°C-29°C-30°C, humidity 65%-63%-62%, and the equipment state feature set of air conditioner No. 1 is running at 26°C for cooling and dehumidifier No. 2 is maintained in low speed mode) and the control strategy feature set (such as strategy A corresponds to the policy element coding feature of "air conditioner No. 1 is adjusted to 24°C for cooling and dehumidifier No. 2 is switched to medium speed", and strategy B corresponds to the policy element coding feature of "turn on the fresh air system, air conditioner No. 3 is maintained at 26°C, and dehumidifier No. 1 is turned off", etc.). Taking the control strategy characteristics of strategy A as an example, the gated loop operator uses this strategy characteristic as the "strategy index vector" (precisely identifying the device direction and parameter requirements of strategy A), and uses the temperature and humidity fluctuation information in the fusion environment feature set as the "environmental sensor vector" (reflecting the current temperature and humidity change trend), and the real-time operation parameters of the equipment as the "equipment state vector" (such as the current operation data of air conditioner No. 1 at 26℃); first calculate the element-by-element product of the strategy index vector and the environmental sensor vector to obtain the "first interaction feature" (characterizing the preliminary correlation strength between the cooling demand of strategy A and the current warming trend); then, based on the "Cooling, Dehumidifying, V The product of the first interaction feature and the multiplicative inverse of the number of control directions is calculated to obtain the "second interaction feature" (balancing the influence weights of different control dimensions); the second interaction feature is mapped to the "first mapping feature" (enhancing the nonlinear expression ability of the feature) through the ReLU nonlinear transformation subfunction built into the gated loop operator; then the element-by-element product of the first mapping feature and the device state vector is calculated to obtain the "third interaction feature" (strengthening the association between strategy A and the current operating state of air conditioner No. 1), and the third interaction feature is added to the strategy index vector element-by-element to generate the "original gated loop feature" (integrating multi-dimensional association information of strategy, environment, and device). After the original gated cycle feature is generated, the server performs standardization processing on it: the average value of all neuron activations in the original gated cycle feature (such as the calculated mean μ = 0.3) and the standard deviation (such as the standard deviation σ = 0.15) are counted; the difference between each element in the original gated cycle feature and the mean μ is used as the "centered feature quantity", and the standard deviation σ is added to the preset minimum value ε (such as 1e-8) and the square root is taken to obtain the "standardization coefficient". Finally, the "gated cycle standardized feature" is obtained by multiplying the centralized feature quantity with the inverse element of the standardization coefficient (eliminating dimensional differences and ensuring the stability of feature distribution).Next, the server inputs the gated cycle normalized features into the fully connected layer (adjusts the feature dimension to the preset size), and processes them through the GELU activation function to obtain the "gated cycle conduction features" (introducing a nonlinear mechanism to improve feature discrimination); then, a residual connection operation is performed on the gated cycle normalized features and the gated cycle conduction features (i.e., element-by-element addition) to obtain the "gated cycle coupling features" (while retaining the original normalized feature information, the conduction features after nonlinear transformation are integrated to avoid the gradient disappearance problem); finally, the gated cycle coupling features are standardized again (repeated mean and standard deviation calculations, centering and coefficient scaling steps), and the final output feature set is the "gated cycle data set corresponding to the control strategy feature set" (where each element corresponds to a quantitative value of the adaptation relationship between a control strategy and the current storage environment. For example, the gated cycle value corresponding to strategy A is higher than that of strategy B, indicating that strategy A is more adapted to the current temperature, humidity and equipment status).

[0031] In an embodiment of the present invention, the strategy generation module includes multiple gated loop branches, the multiple gated loop branches are cascaded, and the multiple gated loop branches include a target gated loop branch; the strategy generation module performs gated loop processing on the fusion environment feature set and the control strategy feature set to obtain a gated loop data set corresponding to the control strategy feature set, including:

[0032] The loop precursor feature corresponding to the target gated loop branch and the fusion environment feature set are used as the loop input feature of the target gated loop branch, and the loop input feature of the target gated loop branch is gated loop processed in the target gated loop branch to obtain the loop response feature of the target gated loop branch; if the target gated loop branch is the first gated loop branch among the multiple gated loop branches, the loop precursor feature corresponding to the target gated loop branch is the control strategy feature set;

[0033] The loop response feature of the target gated loop branch is used as the loop precursor feature corresponding to the subsequent gated loop branch, the loop precursor feature corresponding to the subsequent gated loop branch and the fusion environment feature set are used as the loop input feature of the subsequent gated loop branch, and the loop input feature of the subsequent gated loop branch is gated and processed in the subsequent gated loop branch to obtain the loop response feature of the subsequent gated loop branch, until the loop response feature of the last gated loop branch among the multiple gated loop branches is obtained, and the loop response feature of the last gated loop branch is determined as the gated loop data set corresponding to the control strategy feature set; the subsequent gated loop branch is the next gated loop branch cascaded to the target gated loop branch.

[0034] In an embodiment of the present invention, for example, in the temperature and humidity control scenario of an e-commerce warehouse, when the server acts as the execution subject to process the multi-cascade gated loop branches of the policy generation module, the process is expanded by taking a policy generation module containing three cascaded gated loop branches (branch 1, branch 2, branch 3) as an example: First, for the first gated loop branch (that is, when the target gated loop branch is branch 1), the server uses the loop precursor features of the branch and the fused environment feature set as loop input features. Since branch 1 is the first branch of the cascade sequence, its cyclic precursor feature is the pre-acquired control strategy feature set (this feature set includes the element coding features of all control strategies, such as "Strategy A: Air conditioner No. 1 is adjusted to 24°C for cooling + Dehumidifier No. 2 is switched to medium speed" and "Strategy B: Turn on the fresh air system + Air conditioner No. 3 maintains 26°C + Dehumidifier No. 1 is turned off"); the fused environment feature set is the integrated features of the current warehouse temperature and humidity and time series dynamics (for example, the temperature and humidity sequence sampled every 10 minutes in the shelf area in the past hour is [29°C, 64%RH], [30°C, 63%RH], [31°C, 62%RH]) and the equipment operating status (for example, the log features of Air conditioner No. 2 cooling at 26°C and Dehumidifier No. 1 maintaining low speed mode). After the server sends these two types of features to branch 1, branch 1 processes them internally through a gated recurrent operator: the control strategy feature set is used as the strategy index vector (accurately identifying the core information of each strategy, such as the device direction and parameter requirements), the temperature and humidity fluctuation trends in the fused environmental feature set are decomposed into environmental sensor vectors (reflecting the dynamic laws of the current temperature rise and humidity drop), and the real-time operating parameters of the equipment (such as 26°C cooling for air conditioner No. 2 and low speed for dehumidifier No. 1) are used as the equipment state vector; the original gated recurrent features are generated through multiplication, inverse element weighting, nonlinear transformation (such as ReLU), vector interaction and other steps within the operator, and then through operations such as normalization and residual connection, the cyclic response features of branch 1 are finally output (this feature preliminarily characterizes the matching strength of each control strategy with the current temperature and humidity "instant fluctuations" and "single device operating status". For example, the correlation between the cooling demand of strategy A and the current high temperature of 31°C is preliminarily quantified in the output of branch 1). Next, the server uses the cyclic response features of branch 1 as the cyclic precursor features of the subsequent gated cyclic branch (branch 2), while still using the fused environment feature set as another input of branch 2.The design of branch 2 focuses on the strategy adaptability of the "device collaboration dimension" (such as the temperature and humidity control efficiency when the air conditioner and dehumidifier are linked). Therefore, during processing, the gated loop operator of branch 2 uses the "loop response characteristics of branch 1" as a supplementary strategy index vector (inheriting the preliminary matching information of branch 1), and again combines the "historical data of inter-device linkage" in the fusion environmental feature set (such as the temperature drop rate and humidity control accuracy when air conditioner No. 2 and dehumidifier No. 1 were running simultaneously in the past) as the new device state vector, while retaining the environmental sensor vector of temperature and humidity fluctuations. The interactive calculation, normalization, residual connection and other processes within the gated loop operator are repeated to generate the loop response characteristics of branch 2. (This feature deepens the adaptability of the strategy under the "multi-device collaboration logic". For example, in strategy A, whether the linkage solution of "air conditioner No. 1 set to 24°C + dehumidifier No. 2 at medium speed" can effectively cope with the device collaboration efficiency in the current high temperature and high humidity scenario of 31°C) Subsequently, the cyclic response feature of branch 2 becomes the cyclic precursor feature of the last gated cyclic branch (branch 3), and the cyclic input feature of branch 3 still contains the predecessor and the fused environment feature set. The design of branch 3 focuses on "long-term control effects" (such as the adaptability of temperature and humidity predictions 2 hours after the policy is executed). Therefore, its gated cyclic operator will call the "temperature and humidity time series prediction feature" in the fused environment feature set (based on the current temperature and humidity time series data, use LSTM to predict the temperature and humidity trends in the next 2 hours, such as predicting that the temperature will reach 32°C and the humidity will reach 60% after 1 hour). It uses the "cyclic response feature of branch 2" as the deep strategy index vector (inheriting the matching information of the first two levels of branches) and uses "future temperature and humidity prediction + equipment long-term operation energy efficiency data" as the new equipment state vector; it is processed through the full process of the gated cyclic operator again, and finally outputs the cyclic response feature of branch 3. At this point, the cyclic response characteristics of the last branch 3 are the "gated cycle data set corresponding to the control strategy feature set." Each element in the data set accurately quantifies the comprehensive adaptation relationship between each control strategy (such as Strategy A and Strategy B) and the four-dimensional environment of "current temperature and humidity, single device status, multi-device collaboration, and future temperature and humidity trends." (For example, the gated cycle value of Strategy A is significantly higher than that of Strategy B, indicating that Strategy A is more suitable for the current warehousing scenario in terms of immediate cooling efficiency, equipment collaborative energy efficiency, and long-term temperature and humidity stability.) Through the progressive processing of three-level cascaded gated cycle branches, the server achieves layer-by-layer coupling of strategies and environmental characteristics from "single-dimensional instant matching" to "multi-dimensional deep adaptation." The final output gated cycle data set provides a quantitative basis for the precise selection of subsequent control strategies.

[0035] In an embodiment of the present invention, the loop precursor feature corresponding to the target gated loop branch and the fusion environment feature set are used as the loop input feature of the target gated loop branch, and the loop input feature of the target gated loop branch is gated and processed in the target gated loop branch to obtain the loop response feature of the target gated loop branch. This can be implemented through the following examples.

[0036] In the target gated cycle branch, gated cycle processing is performed on the cycle precursor feature and the fused environment feature set to obtain an original gated cycle feature;

[0037] Normalizing the original gated cycle features to obtain standardized features;

[0038] The standardized features are processed through a fully connected layer and an activation function to obtain a logical conduction feature, the standardized features and the logical conduction features are fused to perform a residual connection operation to obtain a gated loop environment coupling feature, and the gated loop environment coupling feature is standardized to obtain a loop response feature of the target gated loop branch.

[0039] In an embodiment of the present invention, for example, in an e-commerce warehouse temperature and humidity control scenario, when the server processes the target gated loop branch (taking the first branch of the cascade sequence as an example), it first combines the loop precursor feature of the branch (i.e., the control strategy feature set, including the element coding features of strategies such as "Strategy A: Air conditioner No. 1 is adjusted to 24°C for cooling + Dehumidifier No. 2 is switched to medium speed" and "Strategy B: Turn on the fresh air system + Air conditioner No. 3 is maintained at 26°C + Dehumidifier No. 1 is turned off") with the fusion environment feature set (the temperature and humidity time series of the current shelf area in the past 30 minutes is [30°C, 65%RH], [31°C, 63%RH], and Air conditioner No. 2 is operated at 26°C for cooling). The target branch uses the device state feature set of air conditioner 1 and dehumidifier 2 to maintain low speed mode) as the loop input feature to drive the gated loop processing flow in the target branch: the first step is to generate the original gated loop feature: the gated loop operator of the target branch uses the "control strategy feature set" as the strategy index vector (accurately carrying the core information of strategy A, such as "device pointing (air conditioner 1, dehumidifier 2)" and "parameter requirements (24℃ cooling, medium speed)"), and decomposes the "temperature and humidity fluctuation trend (30-31℃ heating, 65-63%RH dehumidification)" in the fusion environment feature set into the environmental sensing vector, and extracts "the current 26℃ cooling, Dehumidifier No. 1 is running at low speed" as the device state vector; the operator first calculates "strategy index vector × environmental sensor vector" to obtain the first interaction feature (characterizing the initial correlation strength between the cooling demand of strategy A and the current high temperature of 31°C); then, based on the number of three types of control directions covered by the fusion environmental feature set, "Cooling (cooling), Dehumidifying (dehumidification), Ventilation (ventilation)", the second interaction feature is calculated as "the multiplicative inverse element of the first interaction feature × the number of control directions (1 / 3)" to balance the influence weights of the three types of control dimensions; through the operator's built-in ReLU nonlinear transformation The sub-function maps the second interaction feature to the first mapping feature (enhancing the nonlinear expression ability of the feature); then the "first mapping feature × device state vector" is calculated to obtain the third interaction feature (strengthening the association between strategy A and the current operating status of air conditioner No. 2 and dehumidifier No. 1), and the "third interaction feature + strategy index vector" is added element by element to generate the original gated loop feature (integrating the three-dimensional correlation information of "strategy direction, environmental fluctuation, and device status". For example, strategy A reflects the preliminary quantitative values ​​of "theoretical cooling efficiency of 24℃ refrigeration on high temperature of 31℃" and "dehumidifier medium speed on 63% RH humidity" in this feature).In the second step, the standardized features are obtained by standardization: the server counts the average value (such as the calculated mean μ = 0.4) and standard deviation (such as standard deviation σ = 0.2) of all neuron activations in the original gated loop feature; the difference between each element in the original gated loop feature and the mean μ is used as the centralized feature, and then the standardization coefficient is obtained by "adding the standard deviation σ to the preset minimum value ε (such as 1e-8) and taking the square root"; finally, the standardized feature is obtained by element-by-element calculation of "centralized feature ÷ standardization coefficient" (eliminating the dimensional difference of the original feature and ensuring the stability of the distribution of features in different dimensions, for example, the magnitude of the "cooling efficiency correlation value" and the "dehumidification potential correlation value" in the corresponding features of strategy A are unified). The third step is to generate logical conduction features and residual connections: the server inputs the standardized features into the fully connected layer (adjusts the feature dimension to the preset size, such as compressing from 128 dimensions to 64 dimensions), and processes them through the GELU activation function to obtain logical conduction features (introducing a nonlinear mechanism to enhance the feature's discriminative power for complex "strategy-environment" associations. For example, after GELU activation, the adaptability of strategy A's "24°C cooling" to a high temperature of 31°C is more accurately highlighted). Subsequently, a residual connection operation (i.e., element-by-element addition) is performed on the "standardized features" and the "logical conduction features" to obtain the gated recurrent environment coupling features (retaining the original information of the standardized features while integrating the conduction features after nonlinear transformation to avoid the gradient vanishing problem. For example, after the residual connection, the "device pointing association value" of strategy A retains the basic matching degree and superimposes the "potential gain of coordinated cooling between air conditioners 1 and 2" mined by the fully connected layer). The fourth step is to normalize again to obtain the loop response characteristics: the server repeats the standardization process of "mean and standard deviation calculation, centering, and coefficient scaling" for the gated loop-environment coupling characteristics, and finally outputs the loop response characteristics of the target gated loop branch. (This characteristic provides subsequent cascade branches with deep coupling information between "Strategy A and the current temperature, humidity, and single device status." For example, the loop response characteristic value of Strategy A is higher than that of Strategy B, indicating that in the first branch processing, Strategy A is more adaptable to "single-device control in immediate high-temperature scenarios.") Through the four-layer processing logic of the target gated loop branch, the server realizes the progressive coupling of "strategy-environment" characteristics from "raw association" to "normalization-nonlinear enhancement-residual fusion-renormalization," laying the foundation for deep adaptation analysis of multiple cascade branches.

[0040] In the embodiment of the present invention, in the target gated loop branch, gated loop processing is performed on the loop precursor feature and the fused environment feature set to obtain the original gated loop feature, which can be implemented through the following example.

[0041] In the target gated loop branch, the loop predecessor feature is used as a policy index vector in a gated loop operator, the fused environment feature set is used as an environment sensing vector in the gated loop operator, and the fused environment feature set is used as a device state vector in the gated loop operator;

[0042] Determining a multiplication result of the strategy index vector and the environment sensing vector as a first interaction feature through the gated loop operator;

[0043] Obtaining the number of control directions corresponding to the fusion environment feature set, and determining a second interaction feature as a result of multiplying the first interaction feature and the multiplication inverse of the number of control directions;

[0044] Converting the second interaction feature into a first mapping feature based on a nonlinear transformation subfunction in the gated loop operator;

[0045] A multiplication result of the first mapping feature and the device state vector is determined as a third interaction feature, and an addition result of the third interaction feature and the policy index vector is determined as an original gated loop feature.

[0046] In an embodiment of the present invention, for example, in the e-commerce warehouse temperature and humidity control scenario, when the server processes the target gated loop branch (taking the first branch of the cascade sequence as an example) to generate the original gated loop feature, the process is as follows: First, the server assigns vector roles to the loop precursor feature and the fusion environment feature set: the loop precursor feature (that is, the element coding feature of "Strategy A: Air conditioner No. 1 is adjusted to 24°C cooling + dehumidifier No. 2 is switched to medium speed" in the control strategy feature set) is used as the strategy index vector of the gated loop operator, which accurately carries the "device pointing (air conditioner No. 1, Dehumidifier No. 2)", "Action instructions (cooling, wind speed adjustment)", "Parameter settings (24℃, medium speed)" and other core information; at the same time, the fusion environment feature set (the fluctuation trend encoding of the temperature and humidity time series of the current shelf area in the past 30 minutes is [30℃, 65%RH]-[31℃, 63%RH], and the device status encoding of air conditioner No. 2 cooling at 26℃ and dehumidifier No. 1 maintaining low speed mode) is synchronized as the operator's environmental sensor vector (focusing on the dynamic change of temperature and humidity over time) and the device state vector (characterizing the real-time operating parameters and collaborative relationship of existing equipment). Next, the server drives the gated loop operator to calculate the first interaction feature: through the element-by-element product operation of the strategy index vector and the environmental sensor vector, the initial correlation strength between strategy A's "cooling and dehumidification demand" and the current "30-31°C warming, 65-63% RH dehumidification" environmental trend is quantified. For example, the cooling demand of "air conditioner No. 1 adjusted to 24°C cooling" in strategy A is multiplied by the time series feature of "31°C high temperature" in the environmental sensor vector to obtain the basic value of the matching degree between the strategy and the environment in this dimension; similarly, the correlation between "dehumidifier No. 2 medium speed" and "63% RH humidity" is also incorporated into the first interaction feature through multiplication, realizing the preliminary aggregation of multi-dimensional correlations. Subsequently, the server obtains the number of control directions corresponding to the fusion environment feature set (the warehouse presets three types of control directions: "Cooling (active cooling), Dehumidifying (active dehumidification), and Ventilation (natural ventilation)", which is 3 in number), and calculates the multiplicative inverse of the number of control directions (i.e. 1 / 3); the first interaction feature is multiplied element-by-element by the multiplicative inverse to obtain the second interaction feature. This step uses weight balancing to avoid the characteristics of a certain type of control direction (such as Cooling) from excessively affecting subsequent calculations due to the dominance of numerical magnitude or physical meaning, ensuring that the three policy dimensions of "cooling, dehumidification, and ventilation" contribute evenly to the final feature.Based on the built-in nonlinear transformation sub-function of the gated loop operator (such as the ReLU activation function), the server inputs the second interactive feature into the sub-function for mapping to generate the first mapping feature. The nonlinear characteristics of ReLU can enhance the feature expression of strategy A in the "cooling adaptability" dimension (such as highlighting the theoretical energy efficiency advantage of 24°C cooling over 31°C high temperature), while suppressing dimensions that are weakly associated with the current environment (such as the redundant association between the dehumidification instruction and the low humidity of 63% RH in strategy A), thereby improving the feature's discriminative power for the core "strategy-environment" association. Finally, the server calculates the third interaction feature and the original gated loop feature: first, the first mapping feature is multiplied element-by-element by the device state vector (the encoding feature of air conditioner No. 2 cooling at 26°C and dehumidifier No. 1 running at low speed) to obtain the correlation strength between strategy A and the real-time status of the existing equipment (such as the synergistic cooling potential of "air conditioner No. 1 adjusted to 24°C" and "air conditioner No. 2 currently 26°C", and the wind speed level complementarity of "dehumidifier No. 2 medium speed" and "dehumidifier No. 1 low speed"); then the third interaction feature is added element-by-element to the strategy index vector, integrating the "device parameter direction of the strategy itself" and the "dynamic association between environment-device-strategy", and finally generating the original gated loop feature. This feature fully characterizes the initial adaptation information of strategy A in the three dimensions of "immediate temperature and humidity fluctuations, existing equipment status, and balance of three types of control directions", providing basic data support for subsequent standardization and residual connection. Through the hierarchical processing of this series of gated loop operators, the server realizes the transformation of the multi-dimensional characteristics of "strategy-environment-device" from "independent existence" to "deep interactive aggregation", laying a solid foundation for the underlying logic to output accurate loop response characteristics for the target gated loop branch.

[0047] In the embodiment of the present invention, the normalization process is performed on the original gated cycle feature to obtain the normalized feature, which can be implemented through the following example.

[0048] Obtaining neuron activation amounts in the original gated cycle features, and determining an average value and a standard deviation of the neuron activation amounts;

[0049] Determine the difference between the original gated cycle feature and the average value as a centralized feature value;

[0050] Obtaining a preset normalization parameter, and taking a square root of the sum of the standard deviation and the preset normalization parameter to obtain a normalization coefficient;

[0051] A multiplication result between the centered feature quantity and the multiplication inverse element of the normalization coefficient is determined as a normalization feature.

[0052] In an embodiment of the present invention, in the e-commerce warehouse temperature and humidity control scenario, when the server performs standardization processing on the original gated cycle features output by the target gated cycle branch, taking the original gated cycle features corresponding to strategy A (carrying the correlation information of "air conditioner No. 1 adjusting 24℃ cooling + dehumidifier No. 2 medium speed" and the current high temperature of 31℃, 63% RH humidity and real-time status of the equipment) as an example, the process is as follows: First, the server obtains the neuron activation amount in the original gated cycle feature. The feature exists in the form of a vector, and each element corresponds to the "cooling adaptability (such as the theoretical energy efficiency of strategy A 24℃ cooling to 31℃ high temperature)" "dehumidification adaptability (the control potential of dehumidifier No. 2 medium speed to 63% RH humidity)" "equipment synergy (cooling synergy between air conditioner No. 1 and the currently running air conditioner No. 2)" "long-term stability (adaptability of temperature and humidity prediction 2 hours after strategy execution)" and other dimensions (assuming the vector is [0.5, 0.3, 0.7, 0.4]). The server traverses all elements of the vector and calculates the average value ((0.5+0.3+0.7+0.4)÷4=0.475) and the standard deviation (first find the sum of the squares of the differences between each element and the mean: (0.5-0.475)) 2 +(0.3-0.475) 2 +(0.7-0.475) 2 +(0.4-0.475) 2 =0.000625+0.030625+0.050625+0.005625=0.0875; then divide by the number of elements 4 to get the variance 0.021875, and the standard deviation is ). Next, the server calculates the centralized feature quantity, subtracts each element of the original gated loop feature from the average value element by element, and obtains the offset of the correlation strength of each dimension relative to the overall mean (such as [0.5-0.475, 0.3-0.475, 0.7-0.475, 0.4-0.475] = [0.025, -0.175, 0.225, -0.075]), eliminating the "mean shift" problem of feature distribution. Then, the server obtains the preset normalization parameter (such as ε = 1×10 -8 , to avoid meaningless calculation when the standard deviation is 0), add the standard deviation to the parameter (0.1479+1×10 -8 ≈0.1479) and then take the square root to get the standardized coefficient The multiplicative inverse of the normalization coefficient is then calculated (i.e., 1÷0.3846≈2.599). Finally, the server multiplies each element of the centralized feature by the multiplicative inverse of the normalization coefficient element-by-element to obtain the normalized feature (e.g., [0.025×2.599,-0.175×2.599,0.225×2.599,-0.075×2.599]≈[0.0649,-0.4548,0.5848,-0.1949]). This process, through "mean centering and standard deviation scaling," eliminates the dimensional differences and distribution shifts of different dimensions in the original gated recurrent features, ensuring that the correlation strengths of dimensions such as "cooling adaptation" and "dehumidification adaptation" are comparable at the same scale, providing a stable feature input foundation for subsequent fully connected layers and residual connections. Through standardized processing, the server realizes the transformation of the original gated loop features from "multi-dimensional heterogeneous correlation" to "comparable features at the same scale", ensuring the stability and accuracy of the subsequent processing logic of the target gated loop branch.

[0053] In an embodiment of the present invention, the target temperature and humidity control model also includes a decision output unit, which outputs the adaptation probability values ​​corresponding to the multiple control strategies respectively according to the gated cycle data set, and sets the target control strategy among the multiple control strategies for the current temperature and humidity time series data according to the adaptation probability value. It can be implemented through the following examples.

[0054] Performing confidence normalization processing on the gated cycle data set based on the nonlinear transformation function of the decision output unit to obtain adaptation probability values ​​corresponding to the multiple control strategies respectively;

[0055] An adaptation probability value exceeding a preset probability threshold among the multiple adaptation probability values ​​is determined as a target adaptation probability value, and a control strategy corresponding to the target adaptation probability value is strategically deployed with the current temperature and humidity time series data.

[0056] In an embodiment of the present invention, for example, first, an adaptation probability value is generated based on a nonlinear transformation function: a Softmax nonlinear transformation function is built into the decision output unit, and the server inputs the gated loop data set output by the strategy generation module (such as the adaptation quantized value vector [1.5, 0.9, 0.6] of the control strategies such as "Strategy A: Air conditioner No. 1 adjusts to 24°C for cooling + dehumidifier No. 2 at medium speed" and "Strategy B: Fresh air + air conditioner No. 3 maintains 26°C") into the function. Softmax converts the quantized value of each strategy into a probability in the range of 0-1 through exponential normalization calculation (the formula is where x i is the quantitative value of the i-th strategy in the gated cycle dataset). Taking strategy A as an example, if its gated cycle value is 1.5, strategy B is 0.9, and strategy C is 0.6, then e 1.5 ≈4.4817,e 0.9≈2.4596,e 0.6 ≈1.8221, and the sum is 4.4817+2.4596+1.8221=8.7634; the adaptation probability of strategy A is 4.4817÷8.7634≈0.511, strategy B is 2.4596÷8.7634≈0.281, and strategy C is 1.8221÷8.7634≈0.208. Finally, the adaptation probability value set corresponding to each control strategy is obtained (such as {Strategy A: 0.511, Strategy B: 0.281, Strategy C: 0.208}). Next, the target adaptation probability value is selected and the strategy is deployed: the server calls the preset probability threshold (such as 0.3), traverses all adaptation probability values, and selects the strategies that exceed the threshold (for example, of 0.511 for strategy A and 0.281 for strategy B, only strategy A meets the threshold). The adaptation probability of strategy A is determined as the target adaptation probability value. Subsequently, based on the warehouse's "energy efficiency first + response speed first" priority rule (Strategy A estimates that cooling to 24°C will take 30 minutes and consume 12 kWh; Strategy B estimates that cooling to 26°C will take 50 minutes and consume 18 kWh), Strategy A is determined to be superior in terms of adaptability, response speed, and energy consumption, and is therefore selected as the target control strategy. Finally, the policy deployment is executed: the server converts the instructions of the target control strategy (Strategy A) into device control signals. Through the warehouse IoT platform, it sends a "start cooling mode, set temperature to 24°C" command to Air Conditioner 1 and a "adjust fan speed to medium" command to Dehumidifier 2. Simultaneously, the policy deployment record is associated with the current temperature and humidity time series data (e.g., the time period data for the shelf area at 31°C and 63% RH) and stored in the warehouse management system, binding the control strategy to the current environment. Through probabilistic normalization and priority screening within the decision output unit, the server achieves precise conversion from "strategy-environment adaptation quantitative value" to "executable control instructions." For example, when the storage temperature suddenly rises to 31°C in the afternoon in summer, Strategy A is calculated to have the highest adaptation probability and the best energy efficiency, and it will promptly drive the equipment to cool down and dehumidify, ensuring a stable cargo storage environment while reducing operational energy consumption.

[0057] In the embodiment of the present invention, the acquisition of control strategy feature sets corresponding to multiple control strategies may be implemented through the following examples.

[0058] Acquire multiple control strategies, deconstruct the multiple control strategies respectively, and obtain the strategy elements corresponding to the multiple control strategies;

[0059] Encoding multiple strategy elements through the feature extraction module in the target temperature and humidity control model to obtain strategy coding feature sets corresponding to the multiple strategy elements respectively;

[0060] Perform weighted aggregation on multiple strategy encoding feature sets respectively to obtain control strategy features corresponding to multiple control strategies;

[0061] The features of multiple control strategies are integrated to obtain control strategy feature sets corresponding to the multiple control strategies.

[0062] In an embodiment of the present invention, for example, in an e-commerce warehouse temperature and humidity control scenario, when the server executes the process of "obtaining a control strategy feature set corresponding to multiple control strategies", it first retrieves multiple preset control strategies from the warehouse strategy database (such as strategy 1: "Air conditioner No. 1 is adjusted to 24°C for cooling, and the wind speed of dehumidifier No. 2 is switched to medium speed", strategy 2: "Turn on the fresh air system, keep air conditioner No. 3 running at 26°C, and turn off dehumidifier No. 1"). The server deconstructs the elements of each strategy: Strategy 1 is decomposed into three types of strategy elements: "Equipment (Air conditioner No. 1, Dehumidifier No. 2)", "Action (cooling, wind speed adjustment)", and "Parameter (24°C, medium speed)"; Strategy 2 is decomposed into three types of strategy elements: "Equipment (fresh air system, Air conditioner No. 3, Dehumidifier No. 1)", "Action (turn on, keep, turn off)", and "Parameter (26°C)". Next, the server calls the feature extraction module of the target temperature and humidity control model to encode the strategy elements: for "equipment elements", the feature extraction module generates the equipment encoding vector through the embedding layer associated with the storage equipment ledger (such as air conditioner No. 1 is mapped to the vector [0.2, 0.5, 0.1], and dehumidifier No. 2 is mapped to [0.3, 0.4, 0.2]); for "action elements", the action encoding vector is generated through the pre-trained action instruction dictionary embedding layer (such as "cooling" is mapped to [0.1, 0.6, 0.3], and "wind speed adjustment" is mapped to [0.4, 0.3, 0.5]); for "parameter elements", the parameter encoding vector is generated through numerical normalization and embedding layer (such as "24℃" is mapped to [0.24, 0.1, 0.3], and "medium speed" is mapped to [0.5, 0.2, 0.4]). Therefore, the three types of elements of strategy 1 respectively obtain the device coding subset [[0.2, 0.5, 0.1], [0.3, 0.4, 0.2]], the action coding subset [[0.1, 0.6, 0.3], [0.4, 0.3, 0.5]], and the parameter coding subset [[0.24, 0.1, 0.3], [0.5, 0.2, 0.4]], which together constitute the strategy coding feature set of strategy 1; strategy 2 generates the corresponding coding subsets in the same way to form its own strategy coding feature set.Subsequently, the server performs weighted aggregation on the coding feature set of each strategy: based on the weight configuration of "equipment collaboration (0.4), action responsiveness (0.3), and parameter accuracy (0.3)" in the warehousing business, the device coding subset of strategy 1 generates the device aggregate feature by element-by-element weighted average ((0.2×0.4+0.3×0.4,0.5×0.4+0.4×0.4,0.1×0.4+0.2×0.4)=[0.2,0.36,0.12]); the action coding subset is similarly weighted averaged ((0.1×0.3+0.4×0.3,0.6×0.3+0.3 The action aggregation feature is generated by taking the weighted average of the parameter coding subsets ((0.24×0.3+0.5×0.3,0.1×0.3+0.2×0.3,0.3×0.3+0.4×0.3)=[0.222,0.09,0.21]) to generate the parameter aggregation feature. The three types of aggregation features are then concatenated ([0.2,0.36,0.12,0.15,0.27,0.24,0.222,0.09,0.21]) to obtain the control strategy feature of strategy 1. Strategy 2 aggregates the coding subsets of equipment (fresh air system, air conditioner No. 3, dehumidifier No. 1), action (on, hold, off), and parameter (26°C) according to the same weight rule to generate its own control strategy feature. Finally, the server sequentially integrates the control strategy features of all control strategies (e.g., Strategy 1, Strategy 2) to form a control strategy feature set (e.g., [Strategy 1 feature vector, Strategy 2 feature vector, ...]). This provides standardized policy feature input for the gated loop processing of the subsequent strategy generation module. Through the complete process of factor deconstruction, encoding, weighted aggregation, and feature integration, the server transforms warehouse control strategies from "natural language instructions" to "machine-understandable high-dimensional features," laying the foundation for precise matching of current temperature and humidity environments.

[0063] In an embodiment of the present invention, the current temperature and humidity time series data includes multiple temperature and humidity sampling points; the feature extraction module in the target temperature and humidity control model performs feature encoding on the current temperature and humidity time series data and the current equipment operation log to obtain the temperature and humidity time series feature set corresponding to the current temperature and humidity time series data, and the current equipment operation log feature corresponding to the current equipment operation log, which can be implemented through the following example.

[0064] Based on the multiple temperature and humidity sampling points in the current temperature and humidity time series data, multiple temperature and humidity time intervals are generated; the multiple temperature and humidity time intervals are subjected to feature mapping by the feature extraction module in the target temperature and humidity control model to obtain initial time features corresponding to the multiple temperature and humidity time intervals, and the multiple initial time features are injected into the time interval time series identifier to obtain multiple time interval enhanced features; the time interval time series identifier refers to information about the position of a temperature and humidity time interval in the current temperature and humidity time series data; the temperature and humidity time interval is composed of one or more temperature and humidity sampling points among the multiple temperature and humidity sampling points;

[0065] Performing feature coding on the multiple time period enhancement features to obtain time period coding features corresponding to the multiple temperature and humidity time period intervals, and generating a temperature and humidity time series feature set corresponding to the current temperature and humidity time series data based on the time period coding features corresponding to the multiple temperature and humidity time period intervals;

[0066] Deconstructing the current device operation log into multiple log semantic units through the feature extraction module in the target temperature and humidity control model;

[0067] Performing semantic vector mapping on the multiple log semantic units to obtain semantic unit vectors corresponding to the multiple log semantic units, and injecting log position identifiers into the multiple semantic unit vectors to obtain multiple semantic enhancement vectors; the log position identifier refers to information about the position of a log semantic unit in the current device operation log;

[0068] Feature encoding is performed on the multiple semantic enhancement vectors to obtain log semantic unit features corresponding to the multiple log semantic units respectively, and a current device operation log feature corresponding to the current device operation log is generated based on the log semantic unit features corresponding to the multiple log semantic units respectively.

[0069] In an embodiment of the present invention, for example, in an e-commerce warehouse temperature and humidity control scenario, the server drives the feature extraction module of the target temperature and humidity control model to perform feature encoding on the current temperature and humidity time series data and the equipment operation log as follows: Feature encoding of the temperature and humidity time series data (generating a temperature and humidity time series feature set) is performed. The current temperature and humidity time series data is collected from sensors in the warehouse shelf area every 5 minutes, forming a sequence of 4 temperature and humidity sampling points: [28°C, 65% RH], [29°C, 63% RH], [30°C, 62% RH], and [31°C, 61% RH]. The server generates two temperature and humidity time periods according to the rule of "aggregating every 3 sampling points into one period": Period 1: includes the first 3 sampling points (temperature rising from 28-29-30°C, humidity falling from 65-63-62% RH); Period 2: includes the 4th sampling point (temperature 31°C, humidity 61% RH, an independent segment). The feature extraction module performs feature mapping for each time period: For time period 1, the mean temperature (29°C), mean humidity (63.33% RH), and humidity fluctuation variance (1.56) are calculated and encoded into a low-dimensional vector (e.g., [0.2, 0.5, 0.3]) to generate the initial time period features. For time period 2, the temperature (31°C) and humidity (61% RH) (no fluctuation) are calculated and encoded into a vector (e.g., [0.1, 0.4, 0.2]) to generate the initial time period features. The initial time period features are then injected with a time period time series identifier (identifying the position of the time period in the overall time series): Time period 1 corresponds to the "first time period" and the position code [1, 0] is appended to the initial feature to generate the enhanced time period features (e.g., [0.2, 0.5, 0.3, 1, 0]). Time period 2 corresponds to the "second time period" and the position code [0, 1] is appended to generate the enhanced time period features (e.g., [0.1, 0.4, 0.2, 0, 1]). Feature encoding is performed on the enhanced features of the two time periods (using the Transformer encoding layer to learn the temporal correlation between time periods): After encoding the enhanced features of time period 1, they highlight the trend characteristic of "continued temperature increase and continuous humidity decrease"; after encoding the enhanced features of time period 2, they emphasize the anomaly information of "single point sudden increase of 31°C". Finally, the encoded features of the two time periods are sequentially concatenated (e.g., [trend feature vector, anomaly feature vector]) to generate the temperature and humidity time series feature set corresponding to the current temperature and humidity time series data, which fully describes the temporal fluctuations, trend changes, and time-period correlation of the warehouse temperature and humidity. Feature encoding of the equipment operation log (generating the current equipment operation log features) The current equipment operation log is the text recorded by the warehouse equipment management system: "Air conditioner 1, cooling at 26°C from 9:00-10:00; dehumidifier 2, medium speed after 10:00."The server deconstructs the log into six semantic units: Unit 1: "Air Conditioner No. 1" (device identifier); Unit 2: "9:00-10:00" (operating time period); Unit 3: "26°C cooling" (operating parameters); Unit 4: "Dehumidifier No. 2" (device identifier); Unit 5: "After 10:00" (operating time period); Unit 6: "Medium fan speed" (operating parameters). Semantic vector mapping is performed on each semantic unit (using the word embedding layer of the pre-trained BERT model): "Air Conditioner No. 1" is mapped to the vector [0.2, 0.5, 0.3]; "26°C cooling" is mapped to the vector [0.1, 0.4, 0.2]. Semantic unit vectors are generated similarly for the remaining units. Inject log position identifiers (identify the order of units in the log) into the semantic unit vector: Unit 1 (air conditioner No. 1) corresponds to the "first semantic unit", and the position code [1, 0, 0, 0, 0] is spliced ​​after the vector to obtain a semantic enhancement vector (such as [0.2, 0.5, 0.3, 1, 0, 0, 0, 0]); Unit 2 (9:00-10:00) corresponds to the "second semantic unit", and the code [0, 1, 0, 0, 0, 0] is spliced ​​to generate a semantic enhancement vector; and so on, complete the generation of enhanced vectors for all 6 units. Feature encoding is performed on the semantic enhancement vectors (using an LSTM layer to learn semantic associations within the log): The LSTM encoding of the enhancement vectors for Unit 1 (Air Conditioner 1) and Unit 3 (26°C Cooling) strengthens the device-parameter association between "Air Conditioner 1 operating at 26°C Cooling"; the encoding of the enhancement vectors for Unit 4 (Dehumidifier 2) and Unit 6 (Medium Wind Speed) strengthens the device-parameter association between "Dehumidifier 2 operating at Medium Wind Speed"; and the encoding of the enhancement vectors for Unit 2 (9:00-10:00) and Unit 5 (After 10:00) strengthens the time-of-day association between "Air Conditioner 1 operating in the morning and Dehumidifier 2 operating after noon." Finally, the encoded features of the six log semantic units are sequentially aggregated (e.g., weighted average or concatenation) to generate the corresponding device operation log features, which fully capture core information such as the device's start / stop status, operating parameters, and time associations. Through the feature extraction module's dual processes of "time period division-feature mapping-time series coding" and "log deconstruction-semantic mapping-semantic coding", the server achieves accurate digital expression of warehouse temperature and humidity dynamics and equipment operating status, providing fine-grained, multi-dimensional feature input for the subsequent adaptation analysis of the strategy generation module.

[0070] In an embodiment of the present invention, the current temperature and humidity time series data includes multiple temperature and humidity sampling points, the number of the target control strategies is multiple, and the target control strategies include a first control strategy; the method also provides the following implementation methods.

[0071] Based on the multiple temperature and humidity sampling points in the current temperature and humidity time series data, multiple temperature and humidity time intervals are generated, and a temperature and humidity time interval feature set corresponding to the multiple temperature and humidity time intervals and a first strategy baseline feature corresponding to the first control strategy are obtained; the temperature and humidity time interval is composed of one or more temperature and humidity sampling points among the multiple temperature and humidity sampling points;

[0072] The strategy generation module performs gated cycle processing on the temperature and humidity time interval feature set and the first strategy baseline feature, respectively, to obtain a temperature and humidity time interval gated cycle data set corresponding to the first control strategy; each gated cycle value in the temperature and humidity time interval gated cycle data set is used to characterize the adaptation relationship between the first control strategy and a temperature and humidity time interval;

[0073] According to the temperature and humidity time period gated cycle data set, the temperature and humidity time period adaptation probability values ​​corresponding to the multiple temperature and humidity time period intervals are output; according to the temperature and humidity time period adaptation probability values, the target temperature and humidity time period interval associated with the first control strategy is obtained from the multiple temperature and humidity time period intervals, and the first control strategy is set for the target temperature and humidity time period interval.

[0074] In an embodiment of the present invention, for example, in the scenario of refined temperature and humidity control in e-commerce warehouses, when the server executes the time period adaptation process for the multi-objective control strategy, the first control strategy ("air conditioner No. 1 is adjusted to 24°C for cooling, and the wind speed of dehumidifier No. 2 is switched to medium speed") and the current temperature and humidity time series data are used as an example: Step 1: Generate temperature and humidity time intervals and obtain feature sets. The current temperature and humidity time series data is taken from the shelf area sensor and collected every 5 minutes to form a sequence containing 4 temperature and humidity sampling points: [28°C, 65%RH](9:00), [29°C, 63%RH](9:05), [30°C, 62%RH](9:10), [31°C, 61%RH](9:15). The server generates two temperature and humidity time intervals, grouping three consecutive sampling points into a single time period. Time period 1 covers three sampling points from 9:00 AM to 9:10 AM, with a continuous temperature increase from 28°C to 29°C to 30°C and a continuous humidity decrease from 65% to 63% to 62% RH. Time period 2 covers a single sampling point at 9:15 AM, with a sudden temperature increase to 31°C and humidity decrease to 61% RH. The feature extraction module of the target temperature and humidity control model performs a "feature mapping-time series identifier injection-encoding" process for each time period. For Time period 1, statistics such as the mean temperature of 29°C and the mean humidity of 63.33% RH are calculated and encoded into vectors, into which the time series identifier of Time period 1 is injected to generate enhanced features. For Time period 2, the anomalous information for the "single sudden temperature increase of 31°C" is extracted and encoded, and the identifier of Time period 2 is injected. This ultimately generates a temperature and humidity time interval feature set, with Time period 1's features focusing on the "sustained temperature and humidity fluctuation trend" and Time period 2's features focusing on the "single point anomaly." At the same time, the server retrieves the baseline features of the first control strategy from the control strategy library: through the "strategy element deconstruction (deconstructed into elements such as 'Air Conditioner No. 1', '24°C Cooling', 'Dehumidifier No. 2', 'Medium Speed') - encoding (devices, actions, and parameters are embedded in vectors respectively) - weighted aggregation (according to device collaboration, action response, and parameter precise weight integration)" process, the first strategy baseline feature vector containing the control logic is generated. Step 2: Gated loop processing of the strategy generation module. The server drives the strategy generation module, aligns the temperature and humidity period interval feature set (period 1 and period 2 features) with the first strategy benchmark features one by one, and performs gated loop processing: For the period 1 features and the first strategy benchmark features: the gated loop operator of the strategy generation module uses the "trend feature vector" of period 1 as the environmental sensing vector (characterizing the temperature and humidity fluctuation law from 9:00 to 9:10) and the first strategy benchmark features as the strategy index vector (characterizing the control logic of "24°C cooling + medium-speed dehumidification"); through the "vector multiplication (quantifying the initial association between strategy and environment) - inverse element weighting (balancing the control direction weight) - nonlinear mapping (strengthening the core association) - residual integration (retaining the original information)" process, the gated loop value (such as 0.8) of period 1 and the first strategy is calculated to quantitatively characterize the adaptation strength of the two.For the characteristics of time period 2 and the baseline characteristics of the first strategy: Similarly, the "abnormal feature vector" of time period 2 is used as the environmental sensing vector, and the baseline characteristics of the first strategy are used as the strategy index vector; after gated loop processing, the gated loop values ​​of time period 2 and the first strategy are obtained (such as 0.3). Finally, the two gated loop values ​​constitute the temperature and humidity time period interval gated loop data set corresponding to the first control strategy, where each value accurately describes the adaptation relationship between the first strategy and the corresponding time period (time period 1 has a higher degree of adaptation). Step 3: Adaptation probability calculation and time period-level strategy deployment, the server calls the decision output unit, performs Softmax nonlinear transformation on the gated loop data set, and generates the temperature and humidity time period interval adaptation probability value: Time period 1 probability: e. 0.8 ÷(e 0.8 +e 0.3 )≈0.75; probability of period 2: e 0.3 ÷(e 0.8 +e 0.3 )≈0.25. The server presets a probability threshold (such as 0.5) and selects time period 1 (probability 0.75≥0.5) as the target temperature and humidity time period associated with the first control strategy. Subsequently, the server binds the execution instruction of the first control strategy to the time interval corresponding to time period 1 (9:00-9:10): through the warehouse IoT platform, the server sends the "start cooling mode during 9:00-9:10 and set the temperature to 24°C" instruction to air conditioner No. 1, and sends the "adjust the wind speed to medium speed during 9:00-9:10" instruction to dehumidifier No. 2, realizing the precise binding execution of "strategy-time period-device". Through the whole process of "time period feature decomposition-strategy-time period gating adaptation-probability screening-time period deployment", the server realizes the refinement of warehouse temperature and humidity control from "global strategy matching" to "time period level precise intervention". For example, during the period of 9:00-9:10 when the temperature and humidity continue to deteriorate, the first control strategy is deployed first to quickly suppress the temperature rise and humidity drop; while during the period of a single point sudden rise at 9:15, other strategies (such as the linkage of the fresh air system) are triggered to maximize the control efficiency and energy efficiency ratio, and ensure the stability of the cargo storage environment.

[0075] In an embodiment of the present invention, the target temperature and humidity control model is obtained in the following manner and can be implemented through the following examples.

[0076] Inputting the sample temperature and humidity time series data and the equipment operation log associated with the sample temperature and humidity time series data into the original temperature and humidity control model; the original temperature and humidity control model includes an original feature extraction module and an original strategy generation module;

[0077] The original feature extraction module in the original temperature and humidity control model performs feature encoding on the sample temperature and humidity time series data and the equipment operation log to obtain a historical temperature and humidity time series feature set corresponding to the sample temperature and humidity time series data, and a sample device log feature corresponding to the equipment operation log; a feature integration operation is performed on the historical temperature and humidity time series feature set and the sample device log feature to obtain a sample fusion environment feature set;

[0078] Obtaining control strategy instance feature sets corresponding to multiple control strategy instances, performing gated cycle processing on the sample fusion environment feature set and the control strategy instance feature set through the original strategy generation module in the original temperature and humidity control model, to obtain a sample gated cycle data set corresponding to the control strategy instance feature set; each gated cycle value in the sample gated cycle data set is used to characterize the matching degree between the sample temperature and humidity time series data and a control strategy instance;

[0079] The original feature extraction module and the original strategy generation module in the original temperature and humidity control model are optimized according to the sample gated cycle data set to obtain a target temperature and humidity control model, which is used to match the control strategy of storage temperature and humidity.

[0080] In an embodiment of the present invention, exemplarily, in the e-commerce warehouse temperature and humidity control model training scenario, the server executes the training process of the target temperature and humidity control model as follows: first, input the sample data into the original model: the server retrieves the sample temperature and humidity time series data of a certain week in summer from the warehouse history database (such as the temperature and humidity sequence collected every hour from 9:00 to 17:00 every day in the shelf area, showing a fluctuation trend of "9:00 (28°C, 65% RH) - 11:00 (30°C, 63% RH) - 13:00 (32°C, 60% RH) - 15:00 (31°C, 62% RH) - 17:00 (30°C, 64% RH)"), as well as the equipment operation log associated with the time period (such as "Air conditioner No. 1 runs at 26°C from 9:00 to 12:00; dehumidifier No. 2 adjusts the wind speed to medium after 10:00; the fresh air system is turned on from 14:00 to 16:00"). The server inputs these two types of data into the original temperature and humidity control model (with a built-in original feature extraction module and original strategy generation module). Secondly, the original feature extraction module generates a sample fusion environmental feature set: for the sample temperature and humidity time series data, the server follows the rule of "dividing a temperature and humidity time period every 2 hours" to generate three time periods (9:00-11:00, 11:00-13:00, 13:00-15:00, and the remaining 15:00-17:00 is the fourth time period). The original feature extraction module performs "feature mapping-time series identifier injection-encoding" on each time period: the mean temperature of time period 1 (9:00-11:00) is calculated as 29°C, the mean humidity is 64%RH, and the humidity fluctuation variance is 1. After encoding it into a vector, the "first time period" identifier is injected to generate the time period enhancement feature; the remaining three time periods are processed similarly, and the final encoding is obtained as the historical temperature and humidity time series feature set (characterizing the trend, fluctuation, and time period correlation of temperature and humidity over time). For device operation logs, the original feature extraction module deconstructs the logs into nine log semantic units: "Air Conditioner No. 1," "9:00-12:00," "26°C Cooling," "Dehumidifier No. 2," "After 10:00," "Medium Wind Speed," "Fresh Air System," "14:00-16:00," and "On." These are mapped to semantic unit vectors using a pre-trained word embedding model, injecting location identifiers such as "1st Semantic Unit (Air Conditioner No. 1)" and "2nd Semantic Unit (9:00-12:00)" to generate a semantic enhancement vector. After LST M encoding, the sample device log features are aggregated to characterize device start and stop, parameters, and time associations. The server performs "element-by-element weighted concatenation" of the historical temperature and humidity time series feature set with the sample device log features to generate a sample fused environmental feature set (which simultaneously carries high-dimensional features of both temperature and humidity dynamics and the real-time status of the device).Next, the original policy generation module generates a sample gated loop dataset: the server retrieves historically validated control policy instances from the storage policy library (e.g., Policy Instance A: "Adjust air conditioner 1 to 25°C for cooling, and switch dehumidifier 2 to medium speed"; Policy Instance B: "Turn on the fresh air system, maintain air conditioner 3 at 27°C, and turn off dehumidifier 1"). For each policy instance, the "element deconstruction (device, action, parameter) - encoding (embedding layer mapping) - weighted aggregation" process is executed to generate a control policy instance feature set. The original strategy generation module inputs the sample fusion environment feature set and the control strategy instance feature set into a multi-cascade gated loop branch: the first branch uses the control strategy instance feature set as the loop precursor, combines the sample fusion environment feature set to calculate the original gated loop features (to quantify the initial match between the strategy and the environment), and outputs the loop response features after normalization and residual connection; subsequent branches inherit the previous feature iterative processing and ultimately output the sample gated loop dataset (such as a matching degree of 0.85 for strategy instance A and 0.62 for strategy instance B, quantifying the adaptability of the sample temperature and humidity time series to each strategy instance). Finally, the model is tuned to obtain the target model: the server uses the "difference between the sample gated loop dataset and the actual optimal strategy label" as the loss function (such as cross-entropy loss, where the label is the actual effective strategy instance under the manually labeled sample data) and iteratively adjusts the parameters of the original feature extraction module and the original strategy generation module (such as optimizing the feature encoding weights and the coefficients of the gated loop operator) through the backpropagation algorithm. When the loss function converges (e.g., the loss drops to less than 0.001 after 10 consecutive iterations), the server stops training and obtains the target temperature and humidity control model. This model can accurately match the real-time temperature and humidity environment of the warehouse with the control strategy, providing core algorithm support for intelligent temperature and humidity control in actual scenarios. Through the full process of "sample data input - feature encoding integration - strategy instance adaptation quantization - parameter iterative optimization", the server completes the training of the target temperature and humidity control model, ensuring that the model can efficiently identify the adaptive relationship between temperature and humidity trends, equipment status and control strategies in actual warehouse scenarios, and realize intelligent and precise control.

[0081] In an embodiment of the present invention, the original feature extraction module includes a temperature and humidity time series feature extraction unit and a log description feature extraction unit; the original feature extraction module and the original strategy generation module in the original temperature and humidity control model are model tuned according to the sample gated cycle data set to obtain a target temperature and humidity control model, which can be implemented through the following example.

[0082] Obtaining a reference temperature and humidity time series feature corresponding to the sample temperature and humidity time series data, and performing model tuning on the temperature and humidity time series feature extraction unit according to the reference temperature and humidity time series feature and the historical temperature and humidity time series feature set;

[0083] Obtaining a benchmark device log feature corresponding to the device operation log, and performing model tuning on the log description feature extraction unit based on the benchmark device log feature and the sample device log feature;

[0084] Obtaining a standard control strategy template corresponding to the sample temperature and humidity time series data, and generating a strategy matching error value based on the standard control strategy template and the sample gated cycle data set;

[0085] The original strategy generation module in the original temperature and humidity control model is optimized according to the strategy matching error value to obtain a target temperature and humidity control model.

[0086] In an embodiment of the present invention, for example, in the tuning stage of the e-commerce warehouse target temperature and humidity control model, the server performs iterative optimization of the original feature extraction module (including the temperature and humidity time series feature extraction unit and the log description feature extraction unit) and the original strategy generation module as follows: 1. Tuning of the temperature and humidity time series feature extraction unit, the server retrieves the sample temperature and humidity time series data of a certain week in summer from the warehouse history database (such as the temperature and humidity sequence collected every hour from 9:00 to 17:00 in the shelf area: [28°C, 65%RH] (9:00 ), [29℃, 63%RH](10:00), [30℃, 62%RH](11:00), [31℃, 61%RH](12:00), [30℃, 63%RH](13:00)), and obtain the benchmark temperature and humidity time series characteristics corresponding to the data (the ideal feature vector manually labeled by domain experts based on the temperature and humidity fluctuation law and cargo storage threshold, such as [0.8, 0.2, 0.9] to represent the "temperature rise trend intensity", "humidity drop trend intensity" and "time period anomaly" respectively). The temperature and humidity time series feature extraction unit of the original temperature and humidity control model performs "time period division (aggregated into time period 1: 9-11, time period 2: 11-13 every 2 hours) - feature mapping (calculating statistics such as the average temperature of 29°C and the average humidity of 64% RH in time period 1 and encoding them into vectors) - time series identifier injection (adding the "1st time period" position code to the time period 1 feature) - Transformer encoding" on the sample data to generate a historical temperature and humidity time series feature set (for example, the time period 1 feature is [0.7, 0.3, 0.8], and the time period 2 feature is [0.9, 0.1, 0.7]). The server calculates the mean square error (MSE) between the historical temperature and humidity time series feature set and the benchmark temperature and humidity time series feature (for example, the MSE between the time period 1 feature and the benchmark is (0.8-0.7) 2 +(0.2-0.3) 2 +(0.9-0.8) 2=0.03), and adjust the neural network weights of the temperature and humidity time series feature extraction unit through the back-propagation algorithm (such as optimizing the time window parameters of the time period aggregation rule and the weight matrix of the feature map embedding layer), minimize the feature encoding error, and complete the model tuning of this unit. 2. To optimize the log description feature extraction unit, the server retrieves the device operation log associated with the sample temperature and humidity time series data (text record: "Air conditioner No. 1 runs at 26°C from 9:00 to 11:00; Dehumidifier No. 2 adjusts the wind speed to medium after 10:00") and obtains the corresponding benchmark device log features of the log (core information vectors annotated by equipment engineers, such as [0.9, 0.1, 0.8] to represent "Air conditioner No. 1 cooling effectiveness", "Dehumidifier No. 2 wind speed adaptability", and "Device time period coordination" respectively). The log description feature extraction unit of the original feature extraction module performs "semantic unit deconstruction (deconstruction into 6 semantic units: 'Air Conditioner No. 1', '9:00-11:00', '26°C cooling', 'Dehumidifier No. 2', 'After 10:00', and 'Medium wind speed') on the log - semantic vector mapping (using the pre-trained BERT model to generate the word vector for each unit, such as 'Air Conditioner No. 1' is mapped to [0.2, 0.5, 0.3]) - position identifier injection (adding the "first semantic unit" position code [1, 0, 0, 0, 0] to the 'Air Conditioner No. 1' unit) - LSTM encoding (learning semantic associations within the log)" to generate sample device log features (for example, the vector after aggregation is [0.8, 0.2, 0.7]). The server calculates the cosine similarity error between the sample device log features and the reference device log features (for example, the similarity is 0.8×0.9+0.2×0.1+0.7×0.8=1.3, and the normalized error is ) Through backpropagation, the weights of the word embedding layer and the recurrent kernel parameters of the LSTM layer of the log description feature extraction unit are adjusted to optimize the encoding accuracy of log semantics and core operation and maintenance information, completing the unit tuning. 3. Tuning the original policy generation module: The server obtains the standard control policy template corresponding to the sample temperature and humidity time series data (the actual optimal policy annotated by the warehouse operations team, such as "air conditioner No. 1 is adjusted to 25°C for cooling, and dehumidifier No. 2 has a medium wind speed") and generates the policy baseline features of this template through the feature extraction module (after "element deconstruction - encoding - weighted aggregation", the vector is [0.9, 0.1, 0.8]). The original strategy generation module performs gated loop processing on the optimized sample fusion environment feature set (integrated from temperature and humidity time series and equipment log features) and the control strategy instance feature set (including features of the standard strategy template): the first branch uses the control strategy instance feature set as the loop precursor, combines the sample fusion environment feature set to calculate the original gated loop feature, and outputs the loop response feature after standardization and residual connection; the subsequent branches iteratively process and finally generate the sample gated loop data set (for example, the matching degree prediction value corresponding to the standard strategy template is 0.7). The server calculates the strategy matching error value (using cross entropy loss: L = -∑(y i logp i )), where y i is the label 1 of the standard policy template, p i The model predicts a matching degree of 0.7), resulting in a loss value of L = -1 × log0.7 ≈ 0.357. Backpropagation is used to adjust the gated loop branch weights and activation function parameters of the fully connected layer of the original policy generation module to minimize the policy matching error and complete module tuning. After the three-stage iterative tuning described above, the server updates all trainable parameters of the original temperature and humidity control model to obtain the target temperature and humidity control model. This model has high precision in temperature and humidity trend encoding, log semantic parsing, and policy adaptation quantification. It can quickly match the optimal control strategy in actual warehousing scenarios to ensure the stability of the cargo storage environment and the optimal energy efficiency of the equipment.

[0087] In the embodiment of the present invention, the generation of a strategy matching error value based on the standard control strategy template and the sample gated cycle data set can be implemented through the following example.

[0088] Outputting sample adaptation probability values ​​corresponding to the plurality of control strategy instances respectively according to the sample gated cycle data set;

[0089] Performing exponential inverse domain conversion on multiple sample adaptation probability values ​​respectively to obtain multiple matching degree inverse domain quantization values;

[0090] Based on the standard control policy template and the sample adaptation probability value, determining the policy matching identification values ​​corresponding to the multiple control policy instances respectively; the policy matching identification value includes a matching identification value or an unmatched identification value, the matching identification value is used to indicate that the standard control policy template contains a policy template that matches the control policy instance corresponding to the matching identification value, and the unmatched identification value is used to indicate that the standard control policy template does not contain a policy template that matches the control policy instance corresponding to the unmatched identification value;

[0091] A strategy matching error value is generated according to the strategy matching identification values ​​respectively corresponding to the multiple control strategy instances and the multiple matching degree inverse domain quantization values.

[0092] In an embodiment of the present invention, for example, in an e-commerce warehouse model tuning scenario, when the server processes the strategy matching error value generation process, taking the sample temperature and humidity time series data (temperature and humidity sequence of the shelf area 9:00-11:00 [28-29-30℃, 65-63-62%RH]), the standard control strategy template ("air conditioner No. 1 is adjusted to 25℃ for cooling, and the wind speed of dehumidifier No. 2 is kept at medium speed") and the control strategy instances A (same as the standard template), B ("turn on the fresh air system + air conditioner No. 3 is kept at 27℃"), and C ("turn off air conditioner No. 1 + turn on dehumidifier No. 1 high-speed mode") as examples, the following steps are performed: Step 1: Output the sample adaptation probability value, the server calls the decision output unit of the original strategy generation module, performs Softmax nonlinear transformation on the sample gated loop data set obtained by gated loop processing (including the adaptation quantization values ​​[1.6, 1.2, 0.8] of strategies A, B, and C), and calculates the sample adaptation probability of each strategy instance: Strategy A: e 1.6 ÷(e 1.6 +e 1.2 +e 0.8 )≈4.953÷(4.953+3.320+2.225)≈0.45; Strategy B: e 1.2 ÷10.498≈0.32; Strategy C: e 0.8 ÷10.498≈0.21; the final sample adaptation probability value set {A:0.45, B:0.32, C:0.21} is generated, and the quantitative model's confidence in the matching of each strategy with the sample temperature and humidity environment. Step 2: Exponential inverse domain conversion is used to obtain the inverse domain quantization value of the matching degree. The server performs exponential inverse domain conversion on each sample adaptation probability value (the conversion formula is Strengthen the error sensitivity of low-probability strategies through inverse domain mapping): Strategy A: 1 / e 0.45 ≈0.637; Strategy B: 1 / e 0.32 ≈0.726; Strategy C: 1 / e 0.21≈0.810; the set of inverse domain quantization values ​​of the matching degree {A: 0.637, B: 0.726, C: 0.810} is obtained, so that the error contribution of the low-probability strategy is easier to capture in the subsequent loss calculation. Step 3: Determine the strategy matching identification value. The server compares the semantic consistency of the standard control strategy template ("air conditioner No. 1 adjusts to 25℃ cooling + dehumidifier No. 2 wind speed medium") with each control strategy instance: Strategy A is completely consistent with the standard template equipment, actions, and parameters - marked as a matching identification value (1); Strategy B (fresh air + air conditioner No. 3) and Strategy C (turn off air conditioner No. 1 + dehumidifier No. 1 high speed) have no overlap with the standard template equipment combination and action instructions - both are marked as unmatched identification values ​​(0); Generate a strategy matching identification value set {A: 1, B: 0, C: 0} to clarify the supervisory label of "whether the standard template matches the strategy instance". Step 4: Generate a strategy matching error value. The server uses weighted cross entropy loss to calculate the error between the model prediction and the supervisory label. The formula is:, where w i is the strategy importance weight (because strategy A matches the standard template, the weight is set to 2; B and C are set to 1), y i is the policy matching identifier value, q i is the inverse domain quantization value of the matching degree. Substituting the numerical calculations: Strategy A: 2×1×log0.637+(1-1)×log(1-0.637)≈2×(-0.440)+0≈-0.880; Strategy B: 1×0×log0.726+(1-0)×log(1-0.726)≈0+1×(-1.290)≈-1.290; Strategy C: 1×0×log0.810+(1-0)×log(1-0.810)≈0+1×(-1.661)≈-1.661; total loss L = -(-0.880-1.290-1.661) = 3.831 (this loss is minimized during training through backpropagation, guiding the model parameters to iterate towards the "exact matching standard strategy"). Through the hierarchical process of "probability output-inverse domain conversion-labeling-loss calculation", the server accurately quantifies the model's prediction error of the "matching relationship between standard policy template and control instance", provides gradient direction for the parameter iteration of the original policy generation module, ensures that the model learns accurate association rules in the policy adaptation task, and finally outputs a highly generalized target temperature and humidity control model.

[0093] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned intelligent control method for storage temperature and humidity based on deep learning. Figure 2 As shown, Figure 2This is a block diagram of the structure of a computer device 100 provided in an embodiment of the present invention. Computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or exchange, memory 111, processor 112, and communication unit 113 are electrically connected to each other, directly or indirectly. For example, these components can be electrically connected via one or more communication buses or signal lines.

[0094] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. These embodiments have been selected and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the present disclosure and to utilize various embodiments with various modifications as appropriate for the specific application contemplated.

Claims

1. A method for intelligent control of storage temperature and humidity based on deep learning, characterized in that: include: Inputting the current temperature and humidity time series data and the current equipment operation log associated with the current temperature and humidity time series data into the target temperature and humidity control model; The target temperature and humidity control model includes a feature extraction module and a strategy generation module; Performing feature encoding on the current temperature and humidity time series data and the current device operation log by the feature extraction module in the target temperature and humidity control model to obtain a temperature and humidity time series feature set corresponding to the current temperature and humidity time series data, and a current device operation log feature corresponding to the current device operation log; performing a feature integration operation on the temperature and humidity time series feature set and the current device operation log feature to obtain a fused environment feature set; Acquire control strategy feature sets corresponding to multiple control strategies, and perform gated loop processing on the fusion environment feature set and the control strategy feature set through the strategy generation module to obtain a gated loop data set corresponding to the control strategy feature set; Each gated cycle value in the gated cycle data set is used to characterize the adaptation relationship between the current temperature and humidity time series data and a control strategy; The adaptation probability values ​​corresponding to the multiple control strategies are output according to the gated cycle data set, and the target control strategy among the multiple control strategies is set for the current temperature and humidity time series data according to the adaptation probability values.

2. The method according to claim 1, characterized in that The gated loop processing is performed on the fusion environment feature set and the control strategy feature set by the strategy generation module to obtain a gated loop data set corresponding to the control strategy feature set, including: In the strategy generation module, gated loop processing is performed on the fusion environment feature set and the control strategy feature set to obtain original gated loop features; Normalizing the original gated cycle feature to obtain a gated cycle normalized feature; The gated cycle normalization feature is processed through a fully connected layer and an activation function to obtain a gated cycle conduction feature, a residual connection operation is performed on the gated cycle normalization feature and the gated cycle conduction feature to obtain a gated cycle coupling feature, and the gated cycle coupling feature is standardized to obtain a gated cycle data set corresponding to the control strategy feature set.

3. The method according to claim 1, characterized in that The strategy generation module includes multiple gated loop branches, the multiple gated loop branches are cascaded, and the multiple gated loop branches include a target gated loop branch; the strategy generation module performs gated loop processing on the fusion environment feature set and the control strategy feature set to obtain a gated loop data set corresponding to the control strategy feature set, including: The loop precursor feature corresponding to the target gated loop branch and the fusion environment feature set are used as the loop input feature of the target gated loop branch, and the loop input feature of the target gated loop branch is gated loop processed in the target gated loop branch to obtain the loop response feature of the target gated loop branch; if the target gated loop branch is the first gated loop branch among the multiple gated loop branches, the loop precursor feature corresponding to the target gated loop branch is the control strategy feature set; The loop response feature of the target gated loop branch is used as the loop precursor feature corresponding to the subsequent gated loop branch, the loop precursor feature corresponding to the subsequent gated loop branch and the fusion environment feature set are used as the loop input feature of the subsequent gated loop branch, and the loop input feature of the subsequent gated loop branch is gated and processed in the subsequent gated loop branch to obtain the loop response feature of the subsequent gated loop branch, until the loop response feature of the last gated loop branch among the multiple gated loop branches is obtained, and the loop response feature of the last gated loop branch is determined as the gated loop data set corresponding to the control strategy feature set; the subsequent gated loop branch is the next gated loop branch cascaded to the target gated loop branch.

4. The method according to claim 3, characterized in that The method of using the cyclic precursor feature corresponding to the target gated cyclic branch and the fused environment feature set as the cyclic input feature of the target gated cyclic branch, performing gated cyclic processing on the cyclic input feature of the target gated cyclic branch in the target gated cyclic branch, and obtaining the cyclic response feature of the target gated cyclic branch includes: In the target gated loop branch, the loop predecessor feature is used as a policy index vector in a gated loop operator, the fused environment feature set is used as an environment sensing vector in the gated loop operator, and the fused environment feature set is used as a device state vector in the gated loop operator; Determining a multiplication result of the strategy index vector and the environment sensing vector as a first interaction feature through the gated loop operator; Obtaining the number of control directions corresponding to the fusion environment feature set, and determining a second interaction feature as a result of multiplying the first interaction feature and the multiplication inverse of the number of control directions; Converting the second interaction feature into a first mapping feature based on a nonlinear transformation subfunction in the gated loop operator; Determine a multiplication result of the first mapping feature and the device state vector as a third interaction feature, and determine an addition result of the third interaction feature and the policy index vector as an original gated loop feature; Obtaining neuron activation amounts in the original gated cycle features, and determining an average value and a standard deviation of the neuron activation amounts; Determine the difference between the original gated cycle feature and the average value as a centralized feature value; Obtaining a preset normalization parameter, and taking a square root of the sum of the standard deviation and the preset normalization parameter to obtain a normalization coefficient; Determine the result of multiplying the centered feature quantity and the multiplication inverse element of the normalization coefficient as the normalization feature; The standardized features are processed through a fully connected layer and an activation function to obtain a logical conduction feature, the standardized features and the logical conduction features are fused to perform a residual connection operation to obtain a gated loop environment coupling feature, and the gated loop environment coupling feature is standardized to obtain a loop response feature of the target gated loop branch.

5. The method according to claim 1, wherein The target temperature and humidity control model further includes a decision output unit, which outputs adaptation probability values ​​corresponding to the multiple control strategies according to the gated cycle data set, and sets a target control strategy among the multiple control strategies for the current temperature and humidity time series data according to the adaptation probability values, including: Performing confidence normalization processing on the gated cycle data set based on the nonlinear transformation function of the decision output unit to obtain adaptation probability values ​​corresponding to the multiple control strategies respectively; An adaptation probability value exceeding a preset probability threshold among the multiple adaptation probability values ​​is determined as a target adaptation probability value, and a control strategy corresponding to the target adaptation probability value is strategically deployed with the current temperature and humidity time series data.

6. The method according to claim 1, characterized in that The obtaining of the control strategy feature sets corresponding to the multiple control strategies includes: Acquire multiple control strategies, deconstruct the multiple control strategies respectively, and obtain the strategy elements corresponding to the multiple control strategies; Encoding multiple strategy elements through the feature extraction module in the target temperature and humidity control model to obtain strategy coding feature sets corresponding to the multiple strategy elements respectively; Perform weighted aggregation on multiple strategy encoding feature sets respectively to obtain control strategy features corresponding to multiple control strategies; The features of multiple control strategies are integrated to obtain control strategy feature sets corresponding to the multiple control strategies.

7. The method according to claim 1, characterized in that The current temperature and humidity time series data includes multiple temperature and humidity sampling points; the feature extraction module in the target temperature and humidity control model performs feature encoding on the current temperature and humidity time series data and the current device operation log to obtain a temperature and humidity time series feature set corresponding to the current temperature and humidity time series data, and a current device operation log feature corresponding to the current device operation log, including: Based on the multiple temperature and humidity sampling points in the current temperature and humidity time series data, multiple temperature and humidity time intervals are generated; the multiple temperature and humidity time intervals are subjected to feature mapping by the feature extraction module in the target temperature and humidity control model to obtain initial time features corresponding to the multiple temperature and humidity time intervals, and the multiple initial time features are injected into the time interval time series identifier to obtain multiple time interval enhanced features; the time interval time series identifier refers to information about the position of a temperature and humidity time interval in the current temperature and humidity time series data; the temperature and humidity time interval is composed of one or more temperature and humidity sampling points among the multiple temperature and humidity sampling points; Performing feature coding on the multiple time period enhancement features to obtain time period coding features corresponding to the multiple temperature and humidity time period intervals, and generating a temperature and humidity time series feature set corresponding to the current temperature and humidity time series data based on the time period coding features corresponding to the multiple temperature and humidity time period intervals; Deconstructing the current device operation log into multiple log semantic units through the feature extraction module in the target temperature and humidity control model; Performing semantic vector mapping on the multiple log semantic units to obtain semantic unit vectors corresponding to the multiple log semantic units, and injecting log position identifiers into the multiple semantic unit vectors to obtain multiple semantic enhancement vectors; the log position identifier refers to information about the position of a log semantic unit in the current device operation log; Feature encoding is performed on the multiple semantic enhancement vectors to obtain log semantic unit features corresponding to the multiple log semantic units respectively, and a current device operation log feature corresponding to the current device operation log is generated based on the log semantic unit features corresponding to the multiple log semantic units respectively.

8. The method according to claim 1, characterized in that The current temperature and humidity time series data includes multiple temperature and humidity sampling points, the number of the target control strategies is multiple, and the target control strategies include a first control strategy; the method further includes: Based on the multiple temperature and humidity sampling points in the current temperature and humidity time series data, multiple temperature and humidity time intervals are generated, and a temperature and humidity time interval feature set corresponding to the multiple temperature and humidity time intervals and a first strategy baseline feature corresponding to the first control strategy are obtained; the temperature and humidity time interval is composed of one or more temperature and humidity sampling points among the multiple temperature and humidity sampling points; The strategy generation module performs gated cycle processing on the temperature and humidity time interval feature set and the first strategy baseline feature, respectively, to obtain a temperature and humidity time interval gated cycle data set corresponding to the first control strategy; each gated cycle value in the temperature and humidity time interval gated cycle data set is used to characterize the adaptation relationship between the first control strategy and a temperature and humidity time interval; According to the temperature and humidity time period gated cycle data set, the temperature and humidity time period adaptation probability values ​​corresponding to the multiple temperature and humidity time period intervals are output; according to the temperature and humidity time period adaptation probability values, the target temperature and humidity time period interval associated with the first control strategy is obtained from the multiple temperature and humidity time period intervals, and the first control strategy is set for the target temperature and humidity time period interval.

9. The method according to claim 1, characterized in that The target temperature and humidity control model is obtained by: Inputting the sample temperature and humidity time series data and the equipment operation log associated with the sample temperature and humidity time series data into the original temperature and humidity control model; the original temperature and humidity control model includes an original feature extraction module and an original strategy generation module; The original feature extraction module in the original temperature and humidity control model performs feature encoding on the sample temperature and humidity time series data and the equipment operation log to obtain a historical temperature and humidity time series feature set corresponding to the sample temperature and humidity time series data, and a sample device log feature corresponding to the equipment operation log; a feature integration operation is performed on the historical temperature and humidity time series feature set and the sample device log feature to obtain a sample fusion environment feature set; Obtaining control strategy instance feature sets corresponding to multiple control strategy instances, performing gated cycle processing on the sample fusion environment feature set and the control strategy instance feature set through the original strategy generation module in the original temperature and humidity control model, to obtain a sample gated cycle data set corresponding to the control strategy instance feature set; each gated cycle value in the sample gated cycle data set is used to characterize the matching degree between the sample temperature and humidity time series data and a control strategy instance; The original feature extraction module and the original strategy generation module in the original temperature and humidity control model are optimized according to the sample gated cycle data set to obtain a target temperature and humidity control model, which is used to match the control strategy of storage temperature and humidity.

10. A server system, characterized in that: The method comprises a server, wherein the server is configured to execute the method according to any one of claims 1 to 9.

Images