Warehouse temperature and humidity intelligent regulation method and system based on deep learning

Abstract

The application discloses a kind of based on deep learning's warehouse temperature and humidity intelligent regulation and control method and system, it is related to artificial intelligence field, including: current temperature and humidity time series data and associated equipment operation log are first input target temperature and humidity control model containing feature extraction module and strategy generation module;Through feature extraction module, temperature and humidity time series data, equipment operation log are encoded and integrated into fusion environment feature set;After obtaining control strategy feature set, the fusion feature and strategy feature are executed by strategy generation module Gating cycle processing, and the gating cycle dataset representing adaptive relationship is output;Based on the dataset, output adaptive probability, and filter target control strategy.The application realizes the accurate matching of temperature and humidity dynamics and control strategy through deep learning, improves the timeliness and energy efficiency ratio of warehouse temperature and humidity regulation and control, and guarantees the stability of goods storage environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a warehouse temperature and humidity intelligent regulation method and system based on deep learning. BACKGROUND

[0002] Current warehouse temperature and humidity regulation relies on manual experience or fixed rules, and has problems such as response lag, low energy efficiency, and poor adaptability to complex scenes. With the expansion of e-commerce warehouse scale and the increasing sensitivity of goods storage, traditional regulation methods cannot meet the demand for dynamic and accurate control of temperature and humidity. SUMMARY

[0003] The purpose of the present application is to provide a warehouse temperature and humidity intelligent regulation method and system based on deep learning.

[0004] In a first aspect, the present application provides a warehouse temperature and humidity intelligent regulation method based on deep learning, comprising:

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

[0006] Encoding the current temperature and humidity time series data and the current device operation logs by the feature extraction module in the target temperature and humidity regulation model to obtain temperature and humidity time series features corresponding to the current temperature and humidity time series data, and current device operation log features corresponding to the current device operation logs, and performing feature integration operations on the temperature and humidity time series features and the current device operation log features to obtain a fusion environment feature set;

[0007] Obtaining a plurality of regulation strategy feature sets corresponding to a plurality of regulation strategies, and performing gated recurrent processing on the fusion environment feature set and the regulation strategy feature set by the strategy generation module to obtain a gated recurrent data set corresponding to the regulation strategy feature set; each gated recurrent value in the gated recurrent data set is used to represent the adaptive relationship between the current temperature and humidity time series data and a regulation strategy;

[0008] Outputting adaptive probability values corresponding to the plurality of regulation strategies respectively according to the gated recurrent data set, and setting a target regulation strategy in the plurality of regulation strategies for the current temperature and humidity time series data according to the adaptive probability values.

[0009] In a second aspect, the present application provides a server system comprising a server, the server being configured to execute the method of the first aspect.

[0010] Compared with the prior art, the application provides the following beneficial effects: by using the warehouse temperature and humidity intelligent regulation and control method and system based on deep learning, the current temperature and humidity time series data and the associated equipment operation log are input into a target temperature and humidity regulation and control model including a feature extraction module and a strategy generation module; the temperature and humidity time series data and the equipment operation log are encoded and integrated into a fusion environment feature set by the feature extraction module; after obtaining the regulation and control strategy feature set, the fusion feature and the strategy feature are subjected to a gating cycle processing by the strategy generation module, and a gating cycle data set representing an adaptive relationship is output; an adaptive probability is output based on the data set, and a target regulation and control strategy is screened. The application realizes accurate matching of temperature and humidity dynamics and regulation and control strategies through deep learning, improves the timeliness and energy efficiency ratio of warehouse temperature and humidity regulation and control, and guarantees the stability of the goods storage environment. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0012] Figure 1 The step flowchart of the warehouse temperature and humidity intelligent regulation and control method based on deep learning provided by the embodiments of the application is shown in the figure.

[0013] Figure 2 The structural schematic block diagram of the computer device provided by the embodiments of the application is shown in the figure. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the embodiments of the application more clear, the following will combine the drawings in the embodiments of the application to clearly and completely describe the technical solutions in the embodiments of the application. Obviously, the described embodiments are some of the embodiments of the application, not all the embodiments. The components of the embodiments of the application described and shown in the drawings can be arranged and designed in various different configurations.

[0015] The specific embodiments of the application will be described in detail below with reference to the drawings.

[0016] In order to solve the technical problems in the foregoing background art, Figure 1 The flowchart of the warehouse temperature and humidity intelligent regulation and control method based on deep learning provided by the embodiments of the application is shown in the figure, and the following will introduce the warehouse temperature and humidity intelligent regulation and control method based on deep learning in detail.

[0017] Step S201, input the current temperature and humidity time series data and the current device 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 comprises a feature extraction module and a strategy generation module;

[0018] Step S202, the feature extraction module in the target temperature and humidity control model is used for feature coding of the current temperature and humidity time series data and the current device operation log, so as to obtain a temperature and humidity time series feature set corresponding to the current temperature and humidity time series data, a current device operation log feature corresponding to the current device operation log, and perform feature integration operation on the temperature and humidity time series feature set and the current device operation log feature, so as to obtain a fusion environment feature set;

[0019] Step S203, obtain a plurality of control strategy features corresponding to a plurality of control strategies, and perform gated recurrent processing on the fusion environment feature set and the control strategy feature set through the strategy generation module, so as to obtain a gated recurrent data set corresponding to the control strategy feature set; each gated recurrent value in the gated recurrent data set is used to represent the adaptive relationship between the current temperature and humidity time series data and one control strategy;

[0020] Step S204, output the adaptive probability value corresponding to each of the plurality of control strategies according to the gated recurrent data set, and set a target control strategy in the plurality of control strategies for the current temperature and humidity time series data according to the adaptive probability value.

[0021] In the embodiment of the present application, in the actual e-commerce warehouse temperature and humidity control scene, the server as the execution subject first executes the step of "inputting the current temperature and humidity time series data and the current device operation log associated with the current temperature and humidity time series data into a target temperature and humidity control model". It is assumed 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 shelf area, sorting area) every 5 minutes, forms current temperature and humidity time series data (for example, the temperature and humidity sequence of the shelf area in a certain period is [28℃, 65%RH], [29℃, 63%RH], [30℃, 62%RH]……), and the running state of the devices (such as air conditioner No. 1 running at 26℃ for refrigeration from 9:00 to 10:00, dehumidifier No. 2 adjusting the air speed to high gear after 10:00) in the warehouse is recorded as the current device operation log by the device management system. The server synchronously extracts the real-time temperature and humidity time series data and the corresponding device operation log through the Internet of Things data acquisition interface, and inputs them into the pre-trained target temperature and humidity control model. The target temperature and humidity control model is built-in with a feature extraction module and a strategy generation module, the former is responsible for feature analysis and coding of the input data, and the latter is used for generating control strategies based on feature matching.

[0022] Next, the server drives the feature extraction module of the target temperature and humidity control model to carry out feature coding and integration operation on the current temperature and humidity time series data and the current device operation log. For the current temperature and humidity time series data, the feature extraction module first divides it into multiple temperature and humidity period intervals (for example, 30 minutes as a period, and the above-mentioned 5-minute sampling points are aggregated into 6 sampling points to form a period interval), and each temperature and humidity period interval is composed of continuous temperature and humidity sampling points. For each temperature and humidity period interval, the feature extraction module maps the features through a neural network structure such as a multilayer perceptron to generate period initial features (including temperature mean, humidity variance, and other basic statistical features in the period); then, a period time sequence identifier (such as the first period, the second period, indicating its position sequence in the overall temperature and humidity time sequence) is injected into each period initial feature to obtain period enhanced features (at this time, the features not only contain temperature and humidity statistical information, but also associate with the time sequence context). The attention mechanism or Transformer coding is performed on all period enhanced features to obtain period coding features corresponding to each temperature and humidity period interval, and then these period coding features are spliced or weighted fused in time sequence to generate a temperature and humidity time sequence feature set corresponding to the current temperature and humidity time sequence (which completely describes the change trend, fluctuation rule, and correlation of each period of the warehouse temperature and humidity over time). For the current device operation log, the feature extraction module first deconstructs the log text into multiple log semantic units (such as "air conditioner 1" "9:00-10:00" "26℃ cooling" semantic fragments), maps each log semantic unit to a semantic vector through a pre-trained word vector model (such as a BERT variant), and obtains a semantic unit vector; then, a log position identifier (such as the first semantic unit corresponding to the device start information, the second corresponding to the running parameter information) is injected into each semantic unit vector to generate a semantic enhanced vector (which combines semantic information and position context within the log). Sequence encoding (such as LSTM) is performed on all semantic enhanced vectors to obtain log semantic unit features corresponding to each log semantic unit, and finally these features are integrated to generate current device operation log features corresponding to the current device operation log (which contains the start and stop state, running parameters, time correlation, and other key information of the device). After the coding of the above two feature sets is completed, the feature extraction module performs feature integration operation on the temperature and humidity time sequence feature set and the current device operation log feature set through feature splicing, element-by-element weighting, etc. to obtain a fused environment feature set (which contains both the time sequence dynamics of the warehouse temperature and humidity and the real-time state of the device operation, providing a comprehensive environment dimension description for subsequent strategy matching).

[0023] Subsequently, the server performs the steps of "obtaining a plurality of control strategy corresponding control strategy feature set" and "performing gated recurrent processing on the fusion environment feature set and the control strategy feature set by the strategy generation module". The warehouse is pre-configured with multiple sets of temperature and humidity control strategies (such as strategy 1: "start air conditioner 1 to 24℃ refrigeration, dehumidifier 2 to medium speed" and strategy 2: "start fresh air system, air conditioner 3 to keep 26℃, close dehumidifier 1", etc.). The server retrieves these control strategies from the strategy database, first deconstructs each strategy into elements such as device identification, action instruction, parameter setting, etc. (such as strategy 1 is decomposed into "air conditioner 1" "start" "refrigeration" "24℃" "dehumidifier 2" "speed" "medium" and other elements). Then, using the feature extraction module of the target temperature and humidity control model, the elements of each strategy are encoded respectively: the vector representation of the device identification is generated through the embedding layer associated with the device account, the action vector is generated through the embedding layer of the action instruction dictionary, and the parameter vector is generated through the numerical coding module, obtaining the strategy coding feature set corresponding to each strategy element. The strategy coding feature set of each strategy is aggregated with weights (such as device element weight 0.4, action element weight 0.3, parameter element weight 0.3), obtaining the control strategy feature corresponding to a single control strategy; the control strategy features of all control strategies are integrated according to the strategy number, forming a plurality of control strategy corresponding control strategy feature set. Then, the strategy generation module performs gated recurrent processing on the fusion environment feature set and the control strategy feature set. The strategy generation module contains multiple cascaded gated recurrent branches. Taking the first gated recurrent branch as an example, its recurrent predecessor feature is the control strategy feature set, and the server takes this recurrent predecessor feature and the fusion environment feature set as the recurrent input features of this branch. Inside the gated recurrent branch, first, the recurrent predecessor feature (control strategy feature set) is taken as the strategy index vector (used to identify the core features of different control strategies) in the gated recurrent operator, and the fusion environment feature set is taken as the environment sensing vector (reflecting the current warehouse temperature and humidity and device state) and the device state vector (focusing on the real-time parameters of device operation).The element-wise product of the strategy index vector and the environment sensing vector is calculated by the gating recurrent operator calculation strategy to obtain the first interaction feature (which describes the preliminary association between the strategy and the environment). Then, the number of temperature and humidity regulation directions (such as the number of Cooling, Dehumidifying, etc. directions, assuming 3 categories) in the fused environment feature set is calculated, and the product of the first interaction feature and the multiplicative inverse of the number of regulation directions is calculated to obtain the second interaction feature (which balances the influence weight of different regulation directions). Based on the nonlinear transformation sub-function (such as the ReLU activation function) within the gating recurrent operator, the second interaction feature is converted into the first mapping feature (which enhances the nonlinear expression ability of the feature). Then, the element-wise product of the first mapping feature and the device state vector is calculated to obtain the third interaction feature (which strengthens the association between the strategy and the current state of the device), and the third interaction feature is added to the strategy index vector to obtain the original gating recurrent feature (which integrates multi-dimensional association information). The original gating recurrent feature is standardized: the average value and the standard deviation of the neuron activation quantity are calculated, the difference value between the original gating recurrent feature and the average value is calculated as the centralized feature quantity, the standard deviation is added to the preset standardization parameter (such as the minimum value ε = 1e-8) and the square root is calculated to obtain the standardization coefficient, and finally the multiplicative inverse of the centralized feature quantity and the standardization coefficient is multiplied to obtain the standardized feature (which eliminates the dimensional difference and improves the stability of the feature). The standardized feature is processed through a fully connected layer and an activation function (such as GELU) to obtain a logical conduction feature, and then the standardized feature and the logical conduction feature are subjected to a residual connection operation (which retains the original feature information and avoids gradient disappearance) to obtain a gating recurrent environment coupling feature. The feature is standardized again to obtain the cycle response feature of the first gating recurrent branch. The subsequent cascaded gating recurrent branches take the cycle response feature of the previous branch as their own cycle precursor feature, and repeat the above gating recurrent processing procedure until the last gating recurrent branch outputs the cycle response feature, which is the gating recurrent dataset corresponding to the regulation strategy feature set. Each gating recurrent value in the gating recurrent dataset quantitatively represents the adaptation relationship between the current temperature and humidity time series data and a single regulation strategy (the higher the value, the stronger the adaptation degree).

[0024] Finally, the server executes "outputting the adaptive probability values corresponding to the multiple control strategies respectively according to the gating cycle data set, and setting a target control strategy in the multiple control strategies for the current temperature and humidity time series data according to the adaptive probability values". The decision output unit of the target temperature and humidity control model performs confidence normalization processing on the gating cycle data set based on a nonlinear transformation function such as Softmax, and converts each gating cycle value into an adaptive probability value in the interval of 0-1 (for example, the adaptive probability of strategy 1 is 0.85, the adaptive probability of strategy 2 is 0.62, and the adaptive probability of strategy 3 is 0.78). The server predefines a probability threshold (for example, 0.7), filters out the control strategies with adaptive probability values exceeding the threshold (for example, strategy 1 and strategy 3), and determines the adaptive probability values corresponding to these strategies as target adaptive probability values. Subsequently, the server selects the control strategy with the highest target adaptive probability value (for example, the adaptive probability of strategy 1 is 0.85, which is higher than the adaptive probability of strategy 3, which is 0.78) as the target control strategy according to the control priority rules of the warehouse (for example, energy consumption priority, response speed priority, etc.), and deploys the strategy to the device control system of the warehouse: sends the instruction "start refrigeration and set the temperature to 24°C" to air conditioner No. 1, and sends the instruction "adjust the air speed to medium" to dehumidifier No. 2, to realize intelligent control of the warehouse environment corresponding to the current temperature and humidity time series data.

[0025] In the whole process, the server analyzes the time series dynamics of the warehouse temperature and humidity and the device operating state through the feature extraction module of the target temperature and humidity control model, deeply matches the environmental characteristics and control strategies with the help of the gating cycle mechanism of the strategy generation module, and finally accurately selects and deploys the target control strategy through the decision output unit, realizes the intelligent and fine control of the warehouse temperature and humidity, and effectively guarantees the stability of the storage environment of the warehouse goods and the energy efficiency of the device operation. For example, in the scenario of sudden temperature rise in the warehouse in the afternoon in summer, the server quickly identifies the strategy with the highest adaptive degree "start air conditioner No. 1 to 24°C refrigeration + dehumidifier No. 2 medium air speed" through the above process, timely reduces the temperature and humidity in the warehouse, avoids damage to goods due to excessive temperature and humidity, and reduces the invalid operation of the device and the energy cost through accurate strategy matching.

[0026] In a possible implementation, the gating cycle processing of the fusion environment feature set and the control strategy feature set by the strategy generation module to obtain the gating cycle data set corresponding to the control strategy feature set can be implemented through the following examples.

[0027] In the strategy generation module, the fusion environment feature set and the control strategy feature set are subjected to gating cycle processing to obtain original gating cycle features;

[0028] The original gating cycle features are subjected to standardization processing to obtain gating cycle standardized features;

[0029] The gated recurrent normalized feature is processed through a full connection layer and an activation function to obtain a gated recurrent conduction feature, a residual connection operation is performed on the gated recurrent normalized feature and the gated recurrent conduction feature to obtain a gated recurrent coupling feature, and the gated recurrent coupling feature is normalized to obtain a gated recurrent dataset corresponding to the regulation strategy feature set.

[0030] In the embodiments of the present application, for example, when the server performs the gating cycle processing of the fusion environment feature set and the regulation policy feature set by the policy generation module, first, the current warehouse fusion environment feature set (such as the sorting area afternoon temperature and humidity time sequence presents a fluctuation trend of 28℃-29℃-30℃, humidity 65%-63%-62%, and the equipment state feature set of air conditioner No. 1 runs at 26℃ cooling, dehumidifier No. 2 maintains low speed mode) and the regulation policy feature set (such as policy A corresponds to the policy element code feature of "air conditioner No. 1 to 24℃ cooling, dehumidifier No. 2 to medium speed", policy B corresponds to the policy element code feature of "turn on fresh air system, air conditioner No. 3 remains at 26℃, and turn off dehumidifier No. 1", etc.) are driven to interact with the gating cycle operator in the policy generation module. Taking the regulation policy feature of policy A as an example, the gating cycle operator takes the regulation policy feature as a "policy index vector" (accurately identifies the device pointing and parameter requirements of policy A), takes the temperature and humidity fluctuation information in the fusion environment feature set as an "environment sensing vector" (reflects the current temperature and humidity change trend), and takes the real-time running parameters of the equipment as a "device state vector" (such as the current running data of air conditioner No. 1 at 26℃); first, the element-wise product of the policy index vector and the environment sensing vector is calculated to obtain a "first interaction feature" (describing the preliminary correlation strength between the cooling demand of policy A and the current temperature rise trend); then, the product of the first interaction feature and the multiplicative inverse of the number of regulation direction numbers (Cooling, Dehumidifying, Ventilation) in the fusion environment feature set is calculated to obtain a "second interaction feature" (balances the influence weight of different regulation dimensions); through the ReLU nonlinear transformation sub-function built-in the gating cycle operator, the second interaction feature is mapped to a "first mapping feature" (enhances the nonlinear expression ability of the feature); then, the element-wise product of the first mapping feature and the device state vector is calculated to obtain a "third interaction feature" (strengthens the association between policy A and the current running state of air conditioner No. 1), and the third interaction feature and the policy index vector are added element by element to generate an "original gating cycle feature" (integrates multi-dimensional association information of policy, environment, and equipment). After the generation of the original gating cycle feature, the server performs standardization processing: the average value (such as the average value μ=0.3) and the standard deviation (such as the standard deviation σ=0.15) of all neuron activation quantities in the original gating cycle feature are calculated; the difference between each element in the original gating cycle feature and the average value μ is taken as a "centered feature quantity", and the square root of the sum of the standard deviation σ and a preset minimum value ε (such as 1e-8) is taken as a "standardization coefficient"; finally, the product of the centered feature quantity and the multiplicative inverse of the standardization coefficient is obtained to obtain a "gating cycle standardized feature" (eliminates the dimensional difference and ensures the stability of the feature distribution).Subsequently, the server inputs the gated recurrent normalized features into a fully connected layer (adjusts the feature dimension to a preset size), and processes them through a GELU activation function to obtain "gated recurrent conduction features" (introduces a nonlinear mechanism to improve feature discrimination); then performs a residual connection operation (i.e., element-wise addition) on the gated recurrent normalized features and the gated recurrent conduction features to obtain "gated recurrent coupling features" (retains the original normalized feature information, while fusing the conduction features after nonlinear transformation to avoid the problem of gradient disappearance); finally, the gated recurrent coupling features are subjected to standardization processing again (repeat the steps of mean and standard deviation calculation, centering and coefficient scaling), and the final output feature set is the "gated recurrent dataset corresponding to the control strategy feature set" (each element corresponds to a quantitative value of the adaptation relationship between a control strategy and the current warehouse environment, for example, if the gated recurrent value corresponding to strategy A is higher than that of strategy B, it indicates that strategy A is more suitable for the current temperature and humidity and equipment state).

[0031] In the embodiment of the present application, the strategy generation module includes a plurality of gated recurrent branches, which are cascaded, and the plurality of gated recurrent branches include a target gated recurrent branch; the gated recurrent processing of the fusion environment feature set and the control strategy feature set by the strategy generation module to obtain the gated recurrent dataset corresponding to the control strategy feature set includes:

[0032] The cycle predecessor feature corresponding to the target gated recurrent branch and the fusion environment feature set are taken as the cycle input features of the target gated recurrent branch, and the cycle input features of the target gated recurrent branch are subjected to gated recurrent processing in the target gated recurrent branch to obtain the cycle response features of the target gated recurrent branch; if the target gated recurrent branch is the first gated recurrent branch in the plurality of gated recurrent branches, the cycle predecessor feature corresponding to the target gated recurrent branch is the control strategy feature set;

[0033] The cycle response features of the target gated recurrent branch are taken as the cycle predecessor features corresponding to a subsequent gated recurrent branch, and the cycle predecessor features corresponding to the subsequent gated recurrent branch and the fusion environment feature set are taken as the cycle input features of the subsequent gated recurrent branch, and the cycle input features of the subsequent gated recurrent branch are subjected to gated recurrent processing in the subsequent gated recurrent branch to obtain the cycle response features of the subsequent gated recurrent branch, until the cycle response features of the last gated recurrent branch in the plurality of gated recurrent branches are obtained, and the cycle response features of the last gated recurrent branch are determined as the gated recurrent dataset corresponding to the control strategy feature set; the subsequent gated recurrent branch is a gated recurrent branch that is cascaded after the target gated recurrent branch.

[0034] In the embodiments of the present application, for example, in the temperature and humidity regulation scene of e-commerce warehouse, when the server as the execution subject processes the multi-cascade gating loop branch of the strategy generation module, take the strategy generation module containing three cascade gating loop branches (branch 1, branch 2, branch 3) as an example to expand the process: first, for the first gating loop branch (i.e. the target gating loop branch is branch 1), the server takes the loop predecessor feature of this branch and the fusion environment feature set as the loop input feature. Since branch 1 is the first branch in the cascade sequence, its loop predecessor feature is the pre-acquired regulation strategy feature set (which contains the element encoding features of all regulation strategies such as "strategy A: air conditioner No. 1 is adjusted to 24℃ cooling + dehumidifier No. 2 is switched to medium speed" and "strategy B: turn on fresh air system + keep air conditioner No. 3 at 26℃ + turn off dehumidifier No. 1"); and the fusion environment feature set is the integrated feature of the current warehouse temperature and humidity and time sequence dynamics (such as the temperature and humidity sequence sampled every 10 minutes in the shelf area in the last 1 hour is [29℃, 64% RH], [30℃, 63% RH], [31℃, 62% RH]) and device running state (such as air conditioner No. 2 is running at 26℃ cooling, dehumidifier No. 1 is maintaining low speed mode). After the server sends these two types of features into branch 1, branch 1 internally processes through the gating loop operator: takes the regulation strategy feature set as the strategy index vector (accurately identifies the core information such as device pointing and parameter requirement of each strategy), disassembles the temperature and humidity fluctuation trend in the fusion environment feature set into an environment sensing vector (reflecting the dynamic law of current temperature rise and humidity drop), and takes the real-time running parameters of the device (such as 26℃ cooling of air conditioner No. 2 and low speed of dehumidifier No. 1) as the device state vector; through the steps of multiplication, inverse element weighting, nonlinear transformation (such as ReLU), vector interaction, etc. in the operator, the original gating loop feature is generated, and then through standardization, residual connection and other operations, the loop response feature of branch 1 is finally output (which preliminarily depicts the matching strength of each regulation strategy and the "instantaneous fluctuation" of current temperature and humidity and "device single running state", for example, the correlation degree of the cooling demand of strategy A and the current high temperature of 31℃ is preliminarily quantified in the output of branch 1). Then, the server takes the loop response feature of branch 1 as the loop predecessor feature of the subsequent gating loop branch (branch 2), and still takes the fusion environment feature set as the other input of branch 2.The design of branch 2 focuses on the strategy adaptability of the "device coordination dimension" (such as the temperature and humidity regulation efficiency when the air conditioner and dehumidifier are linked), so when processing, the gating loop operator of branch 2 will take the "branch 1 loop response feature" as a supplementary strategy index vector (inherit the preliminary matching information of branch 1), and again combine the "device interlinkage historical data" in the fusion environment feature set (such as the temperature drop rate and humidity control accuracy when air conditioner 2 and dehumidifier 1 run at the same time in the past) as a new device state vector, and keep the temperature and humidity fluctuation environment sensor vector; repeat the interactive calculation, standardization, residual connection and other processes in the gating loop operator to generate the loop response feature of branch 2 (this feature deepens the adaptability of the strategy under the "multi-device coordination logic", for example, whether the linkage scheme of "air conditioner 1 sets 24℃ + dehumidifier 2 medium speed" in strategy A can efficiently cope with the device coordination efficiency in the current 31℃ high temperature and high humidity scene). Subsequently, the loop response feature of branch 2 becomes the loop precursor feature of the last gating loop branch (branch 3), and the loop input feature of branch 3 still contains the precursor and the fusion environment feature set. The design of branch 3 focuses on the "long-term regulation effect" (such as the temperature and humidity prediction adaptability 2 hours after the strategy is executed), so its gating loop operator will call the "temperature and humidity time sequence prediction feature" in the fusion environment feature set (based on the current temperature and humidity time sequence data, through LSTM to predict the temperature and humidity trend in the next 2 hours, such as predicting that the temperature will reach 32℃ and the humidity will be 60% in 1 hour), take the "branch 2 loop response feature" as a deep strategy index vector (inherit the matching information of the previous two branches), and take the "future temperature and humidity prediction + device long-term running energy efficiency data" as a new device state vector; again through the whole process of the gating loop operator, the loop response feature of branch 3 is finally output. At this time, the loop response feature of the last branch 3 is the "gating loop data set corresponding to the regulation strategy feature set", and each element in the data set accurately quantifies the comprehensive adaptation relationship between each regulation strategy (such as strategy A, strategy B) and the four-dimensional environment of "current temperature and humidity, device single-state, multi-device coordination, future temperature and humidity trend" (for example, the gating loop value of strategy A is significantly higher than that of strategy B, indicating that strategy A is more adaptive to the current warehouse scene in terms of immediate cooling efficiency, device coordination energy efficiency, and long-term temperature and humidity stability). Through the progressive processing of the three-level cascading gating loop branches, the server realizes the layer-by-layer coupling of the strategy and the environment feature from "single-dimensional immediate matching" to "multi-dimensional deep adaptation", and finally outputs the gating loop data set, which provides a quantitative basis for the accurate selection of subsequent regulation strategies.

[0035] In the embodiment of the present application, the target gating loop branch corresponding to the loop precursor feature and the fusion environment feature set are taken as the loop input feature of the target gating loop branch, and the loop input feature of the target gating loop branch is processed by gating loop processing in the target gating loop branch to obtain the loop response feature of the target gating loop branch. It can be implemented by the following example.

[0036] In the target gating cycle branch, the cycle precursor feature and the fusion environment feature set are subjected to gating cycle processing to obtain an original gating cycle feature;

[0037] The original gating cycle feature is subjected to standardization processing to obtain a standardized feature;

[0038] The standardized feature is processed through a full connection layer and an activation function to obtain a logical conduction feature. The standardized feature and the logical conduction feature are fused to perform a residual connection operation to obtain a gating cycle environment coupling feature. The gating cycle environment coupling feature is subjected to standardization processing to obtain a cycle response feature of the target gating cycle branch.

[0039] In the embodiment of the application, in the e-commerce warehouse temperature and humidity regulation scenario, for example, when the server processes the target gating loop branch (taking the first branch in the cascade sequence as an example), first, the loop predecessor features of the branch (i.e., the regulation strategy feature set, containing the element code features of strategies such as "Strategy A: air conditioner No. 1 is adjusted to 24 DEG C refrigeration + dehumidifier No. 2 is switched to medium speed" and "Strategy B: turn on fresh air system + air conditioner No. 3 remains at 26 DEG C + turn off dehumidifier No. 1") and the fusion environment feature set (the current shelf area temperature and humidity time sequence in the last 30 minutes is [30 DEG C, 65% RH] and [31 DEG C, 63% RH], and the device state feature set of air conditioner No. 2 operating at 26 DEG C refrigeration and dehumidifier No. 1 maintaining low speed mode) are taken as loop input features to drive the gating loop processing flow in the target branch: first, generate the original gating loop features: the gating loop operator of the target branch takes the "regulation strategy feature set" as the strategy index vector (precisely bearing the core information of "device pointing (air conditioner No. 1, dehumidifier No. 2)" and "parameter requirements (24 DEG C refrigeration, medium speed)" of strategy A), disassembles the "temperature and humidity fluctuation trend (30-31 DEG C temperature rise, 65-63% RH humidity reduction)" in the fusion environment feature set into an environment sensing vector, and extracts "air conditioner No. 2 currently operating at 26 DEG C refrigeration and dehumidifier No. 1 operating at low speed" as a device state vector; the operator first calculates "strategy index vector x environment sensing vector" to obtain the first interaction feature (describing the preliminary correlation strength of the cooling demand of strategy A and the current high temperature of 31 DEG C); then, according to the number of three types of regulation directions "Cooling (cooling)", "Dehumidifying (dehumidifying)", and "Ventilation (ventilation)" included in the fusion environment feature set, calculate the multiplicative inverse (1 / 3) of "the first interaction feature x the number of regulation directions" to obtain the second interaction feature (balance the influence weight of the three types of regulation dimensions); through the ReLU nonlinear transformation sub-function built-in the operator, map the second interaction feature to the first mapping feature (enhance the nonlinear expression ability of the feature); then calculate "the first mapping feature x the device state vector" to obtain the third interaction feature (strengthen the correlation of strategy A with the current operating state of air conditioner No. 2 and dehumidifier No. 1), and add the "third interaction feature + strategy index vector" element by element to generate the original gating loop feature (integrate the three-dimensional correlation information of "strategy pointing", "environment fluctuation", and "device state", such as the preliminary quantitative values of "the theoretical cooling efficiency of 24 DEG C refrigeration on 31 DEG C high temperature" and "the regulation potential of dehumidifier medium speed on 63% RH humidity" of strategy A in the feature).Second step, the standardized features are obtained through standardization: the server calculates the average value (for example, the average value μ=0.4) and the standard deviation (for example, the standard deviation σ=0.2) of the activation amount of all neurons in the original gating cycle feature; the difference between each element in the original gating cycle feature and the average value μ is taken as the centralized feature quantity, and then the standard deviation σ is added to the preset minimum value ε (for example, 1e-8) and the square root is taken to obtain the standardization coefficient; finally, the standardized features are obtained through element-by-element calculation of the centralized feature quantity divided by the standardization coefficient (eliminate the dimensional difference of the original features, ensure the stability of the distribution of features of different dimensions, for example, the order of magnitude of the "cooling efficiency correlation value" and the "dehumidification potential correlation value" corresponding to strategy A is unified). Third step, generate logical conduction features and residual connection: the server inputs the standardized features into a fully connected layer (adjust the feature dimension to a preset size, for example, from 128 dimensions to 64 dimensions), and processes them through a GELU activation function to obtain logical conduction features (introduce a nonlinear mechanism to strengthen the discriminability of features to the "strategy-environment" complex correlation, for example, after GELU activation of "24℃ refrigeration" of strategy A, its adaptability feature to 31℃ high temperature is more accurately highlighted); then, the standardized features and the logical conduction features are subjected to residual connection operation (i.e., element-by-element addition) to obtain gating cycle environment coupling features (retain the original information of the standardized features, while integrating the conduction features after nonlinear transformation, to avoid the problem of gradient disappearance, for example, after residual connection of the "device pointing correlation value" of strategy A, it retains the basic matching degree, and also superimposes the "potential gain of air conditioner No. 1 and air conditioner No. 2 synergistic cooling" information mined by the fully connected layer). Fourth step, standardize again to obtain cycle response features: the server repeats the standardization process of "average value and standard deviation calculation, centralization, and coefficient scaling" on the gating cycle environment coupling features, and finally outputs the cycle response features of the target gating cycle branch (the features provide "strategy A and the current temperature and humidity, and the state of a single device" deep coupling information for the subsequent cascaded branches, for example, if the cycle response feature value of strategy A is higher than that of strategy B, it indicates that in the first branch processing, strategy A is more suitable for "single-device regulation and control adaptability in the immediate high-temperature scenario"). Through the four-layer processing logic of the target gating cycle branch, the server realizes the progressive coupling of "strategy-environment" features from "original correlation" to "standardization-nonlinear reinforcement-residual fusion-restandardization", laying a foundation for deep adaptation analysis of multiple cascaded branches.

[0040] In the embodiment of the present application, the gating cycle processing of the cycle precursor features and the fusion environment feature set in the target gating cycle branch can be implemented through the following examples.

[0041] In the target Gated Recurrent Unit branch, the loop predecessor feature is taken as a policy index vector in a Gated Recurrent Unit, the fused environment feature set is taken as an environment sensing vector in the Gated Recurrent Unit, and the fused environment feature set is taken as a device state vector in the Gated Recurrent Unit;

[0042] A multiplication result of the policy index vector and the environment sensing vector is determined as a first interaction feature through the Gated Recurrent Unit;

[0043] A multiplication result of the first interaction feature and a multiplicative inverse of the number of regulation directions corresponding to the fused environment feature set is determined as a second interaction feature;

[0044] The second interaction feature is converted into a first mapping feature based on a nonlinear transformation sub-function in the Gated Recurrent Unit;

[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 Recurrent Unit feature.

[0046] In the embodiment of the present application, in the e-commerce warehouse temperature and humidity regulation scenario, the server generates the original gated loop feature when processing the target gated loop branch (taking the first branch of the cascade sequence as an example). The process is as follows: first, the server assigns vector roles to the loop predecessor feature and the fused environment feature set: the loop predecessor feature (i.e. the element encoding feature of “Strategy A: air conditioner No. 1 is adjusted to 24℃ cooling + dehumidifier No. 2 is switched to medium speed” in the regulation strategy feature set) is used as the strategy index vector of the gated loop operator, which accurately carries the core information of “device pointing (air conditioner No. 1, dehumidifier No. 2)”, “action instruction (cooling, air speed adjustment)”, “parameter setting (24℃, medium speed)” and the like; at the same time, the fused environment feature set (the fluctuation trend encoding of the current shelf area temperature and humidity time sequence in the last 30 minutes is [30℃, 65%RH]—[31℃, 63%RH], and the device state encoding of air conditioner No. 2 running at 26℃ cooling and dehumidifier No. 1 maintaining low speed mode) is used as the environment sensing vector (focusing on the dynamic change law of temperature and humidity over time) and the device state vector (describing the real-time running parameters and cooperative relationship of existing devices) of the operator. Then, the server drives the gated loop operator to calculate the first interaction feature: through the element-by-element multiplication operation of the strategy index vector and the environment sensing vector, the preliminary correlation strength of the “cooling and dehumidifying demand” of strategy A and the current “30—31℃ temperature rise, 65—63%RH humidity drop” environment trend is quantified, for example, the “cooling demand” of “air conditioner No. 1 adjusted to 24℃ cooling” in strategy A is multiplied by the time sequence feature of “31℃ high temperature” in the environment sensing vector to obtain the basic value of the matching degree of the strategy and the environment in this dimension; similarly, the correlation degree of “dehumidifier No. 2 medium speed” and “63%RH humidity” is also included in the first interaction feature through multiplication to realize the preliminary aggregation of multi-dimensional correlation. Subsequently, the server obtains the number of regulation directions corresponding to the fused environment feature set (the warehouse is preset with “Cooling (active cooling), Dehumidifying (active dehumidifying), Ventilation (natural ventilation)” three types of regulation directions, the number is 3), and calculates the multiplicative inverse of the number of regulation directions (i.e. 1 / 3); the first interaction feature is multiplied by the multiplicative inverse element by element to obtain the second interaction feature, which balances the weight through this step to avoid the feature factor value magnitude or physical meaning of a certain type of regulation direction (such as Cooling) from being dominant and excessively affecting subsequent calculation, ensuring that the contribution of “cooling, dehumidifying, ventilation” three types of strategy dimensions to the final feature is balanced.Based on the nonlinear transformation sub-function (such as ReLU activation function) built in the gated recurrent operator, the server inputs the second interaction feature into the sub-function for mapping to generate the first mapping feature. The nonlinear characteristics of ReLU can strengthen the feature expression of strategy A in the "cooling adaptability" dimension (such as highlighting the theoretical energy efficiency advantage of 24℃ cooling for 31℃ high temperature), while suppressing the dimensions weakly associated with the current environment (such as the redundant association of the dehumidification instruction in strategy A with 63% RH low humidity), and improving the discriminability of the feature to the core association of "strategy-environment". Finally, the server calculates the third interaction feature and the original gated recurrent feature: first, multiply the first mapping feature and the device state vector (the encoding feature of air conditioner No. 2 cooling at 26℃ and dehumidifier No. 1 running at low speed) element by element to obtain the association strength of strategy A and the real-time state of the existing device (such as the synergistic cooling potential of "air conditioner No. 1 at 24℃" and "air conditioner No. 2 currently at 26℃", and the wind speed level complementarity of "dehumidifier No. 2 at medium speed" and "dehumidifier No. 1 at low speed"); then add the third interaction feature and the strategy index vector element by element to integrate the "device parameter pointing of the strategy itself" and the "dynamic association of environment-device-strategy", and finally generate the original gated recurrent feature, which completely describes the initial adaptation information of strategy A in the three dimensions of "immediate temperature and humidity fluctuation", "existing device state" and "three types of regulation direction balance", providing basic data support for subsequent standardization and residual connection. Through this series of hierarchical processing of the gated recurrent operator, the server realizes the transformation of the "strategy-environment-device" multidimensional feature from "independent existence" to "deep interaction aggregation", and builds a solid foundation for the target gated recurrent branch to output accurate recurrent response features.

[0047] In the embodiment of the application, the standardization processing of the original gated recurrent feature to obtain the standardized feature can be implemented by the following examples.

[0048] Obtain the neuron activation quantity in the original gated recurrent feature, and determine the average value and the standard deviation of the neuron activation quantity;

[0049] Determine the difference value between the original gated recurrent feature and the average value as the centralization feature quantity;

[0050] Obtain a preset standardization parameter, square root of the sum of the standard deviation and the preset standardization parameter to obtain a standardization coefficient;

[0051] Determine the multiplication result between the centralization feature quantity and the multiplicative inverse of the standardization coefficient as the standardized feature.

[0052] In the embodiment of the present application, in the e-commerce warehouse temperature and humidity regulation scene, when the server performs standardization processing on the original gating cycle feature output by the target gating cycle branch, taking the original gating cycle feature corresponding to strategy A (carrying the association information of "air conditioner No. 1 24℃ refrigeration + dehumidifier No. 2 medium speed" and the current 31℃ high temperature, 63% RH humidity and the real-time state of the equipment) as an example, the process is as follows: first, the server obtains the neuron activation quantity in the original gating cycle feature, which exists in the form of a vector, and each element corresponds to the association strength value of the dimensions such as "cooling adaptation degree (such as the theoretical energy efficiency of strategy A 24℃ refrigeration to 31℃ high temperature)", "dehumidification adaptation degree (the regulation potential of dehumidifier No. 2 medium speed to 63% RH humidity)", "device synergy degree (cooling synergy of 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 so on (assuming the vector is [0.5, 0.3, 0.7, 0.4]). The server traverses all elements of the vector, calculates the average value ((0.5+0.3+0.7+0.4)÷4=0.475) and the standard deviation (first calculate the sum of squares of the difference between each element and the mean value: (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 ). Then the server calculates the centralized feature quantity, subtracts each element of the original gating cycle feature from the average value element by element to obtain the offset of the association strength of each dimension relative to the overall mean value (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 a preset standardization parameter (such as ε = 1×10 -8 , to avoid meaningless calculation when the standard deviation is 0), adds the standard deviation to the parameter (0.1479+1×10 -8 ≈0.1479) and takes the square root to obtain the standardization coefficient The multiplicative inverse of the standardization coefficient (i.e., 1 ÷ 0.3846 ≈ 2.599) is recalculated. Finally, the server element-by-element multiplies each element of the centralized feature quantity with the multiplicative inverse of the standardization coefficient to obtain the standardized feature (such as [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 eliminates the dimensional differences and distribution offsets of different dimensions in the original gating cycle feature through "mean centering-standard deviation scaling", ensures that the correlation strengths of dimensions such as "cooling adaptation" and "dehumidification adaptation" are comparable on the same scale, and provides a stable feature input basis for subsequent fully connected layers and residual connections. Through standardization processing, the server realizes the transformation of the original gating cycle feature from "multi-dimensional heterogeneous correlation" to "features on the same scale", ensuring the stability and accuracy of the subsequent processing logic of the target gating cycle branch.

[0053] In the embodiment of the present application, the target temperature and humidity regulation model further comprises a decision output unit, which outputs the adaptation probability values corresponding to the plurality of regulation strategies respectively according to the gating cycle data set, and sets a target regulation strategy in the plurality of regulation strategies for the current temperature and humidity time series data according to the adaptation probability values. The following examples can be implemented.

[0054] The gating cycle data set is subjected to confidence standardization processing based on a nonlinear transformation function of the decision output unit, to obtain the adaptation probability values corresponding to the plurality of regulation strategies respectively;

[0055] The adaptation probability values exceeding a preset probability threshold in the plurality of adaptation probability values are determined as target adaptation probability values, and the regulation strategy corresponding to the target adaptation probability values is deployed with the current temperature and humidity time series data.

[0056] In the embodiment of the present application, for example, first, the adaptation probability values are generated based on a nonlinear transformation function: the decision output unit is built-in with a Softmax nonlinear transformation function, and the server inputs the gating cycle data set (such as the adaptation quantitative value vector [1.5, 0.9, 0.6] containing the regulation strategies "strategy A: air conditioner No. 1 cooling to 24℃ + dehumidifier No. 2 medium speed", "strategy B: fresh air + air conditioner No. 3 maintaining 26℃", etc.) output by the strategy generation module into the function. Softmax calculates the probability in the 0-1 interval by exponential normalization, converting the quantitative value of each strategy into a probability (the formula is where x i is the quantitative value of the i-th strategy in the gating cycle data set). Taking strategy A as an example, if its gating cycle value is 1.5, strategy B is 0.9, and strategy C is 0.6, the calculation is e 1.5 ≈4.4817, e 0.9≈2.4596, e 0.6 ≈1.8221, the sum is 4.4817+2.4596+1.8221=8.7634; the adaptation probability of strategy A is 4.4817÷8.7634≈0.511, that of strategy B is 2.4596÷8.7634≈0.281, and that of strategy C is 1.8221÷8.7634≈0.208, and finally the adaptation probability value set corresponding to each regulation strategy is obtained (for example, {strategy A: 0.511, strategy B: 0.281, strategy C: 0.208}). Then, the target adaptation probability value is screened and the strategy is deployed: the server calls a preset probability threshold (for example, 0.3), traverses all the adaptation probability values, and screens the strategies (for example, only strategy A meets among 0.511 of strategy A and 0.281 of strategy B) that exceed the threshold, and determines the adaptation probability of strategy A as the target adaptation probability value. Subsequently, according to the priority rule of "energy efficiency priority + response speed priority" of the warehouse (strategy A is estimated to take 30 minutes and consume 12 kWh to cool to 24℃; strategy B is estimated to take 50 minutes and consume 18 kWh to cool to 26℃), it is confirmed that strategy A is better in adaptation, response speed and energy consumption, and it is used as the target regulation strategy. Finally, the strategy is deployed: the server converts the instructions of the target regulation strategy (strategy A) into device control signals, sends the "start cooling mode and set temperature to 24℃" instruction to air conditioner No. 1 and the "adjust the air speed to medium speed" instruction to dehumidifier No. 2 through the warehouse Internet of Things platform; at the same time, the strategy deployment record is stored in the warehouse management system in association with the current temperature and humidity time series data (for example, the time period data of 31℃ and 63% RH of the shelf area), and the binding execution of the regulation strategy and the current environment is completed. Through the probability standardization and priority screening of the decision output unit, the server realizes the accurate conversion from "strategy-environment adaptation quantitative value" to "executable regulation instruction". For example, when the temperature in the warehouse rises to 31℃ in the afternoon of summer, strategy A has the highest calculated adaptation probability and the best energy efficiency, and timely drives the equipment linkage to cool and dehumidify, which not only ensures the stability of the goods storage environment, but also reduces the operating energy consumption.

[0057] In the embodiments of the present application, the acquisition of the regulation strategy characteristic set corresponding to the plurality of regulation strategies can be implemented by the following examples.

[0058] A plurality of regulation strategies are acquired, and a plurality of regulation strategies are respectively deconstructed to obtain a plurality of strategy elements respectively corresponding to the plurality of regulation strategies;

[0059] The plurality of strategy elements are encoded by the feature extraction module in the target temperature and humidity regulation model to obtain a plurality of strategy encoding feature sets respectively corresponding to the plurality of strategy elements;

[0060] The plurality of strategy encoding feature sets are respectively weighted and aggregated to obtain a plurality of regulation strategy characteristics respectively corresponding to the plurality of regulation strategies;

[0061] The multiple regulation strategy characteristics are integrated to obtain a regulation strategy characteristic set corresponding to the multiple regulation strategies.

[0062] In the embodiment of the present application, in the e-commerce warehouse temperature and humidity regulation scene, when the server executes the process of "obtaining the regulation strategy characteristic set corresponding to the plurality of regulation strategies", it first retrieves a plurality of preset regulation strategies (such as strategy 1: "air conditioner 1 is adjusted to 24℃ for refrigeration, and dehumidifier 2 is switched to medium speed", and strategy 2: "turn on the fresh air system, keep air conditioner 3 running at 26℃, and turn off dehumidifier 1") from the warehouse strategy database. The server deconstructs each strategy into three types of strategy elements: strategy 1 is disassembled into "equipment (air conditioner 1, dehumidifier 2)", "action (refrigeration, speed adjustment)", and "parameter (24℃, medium speed)"; strategy 2 is disassembled into "equipment (fresh air system, air conditioner 3, dehumidifier 1)", "action (turn on, keep, turn off)", and "parameter (26℃)". Then, the server calls the feature extraction module of the target temperature and humidity regulation model to encode the strategy elements: for "equipment elements", the feature extraction module generates equipment encoding vectors through the embedding layer associated with the warehouse equipment account (such as air conditioner 1 is mapped to vector [0.2, 0.5, 0.1], and dehumidifier 2 is mapped to [0.3, 0.4, 0.2]); for "action elements", the action instruction dictionary embedding layer (such as "refrigeration" is mapped to [0.1, 0.6, 0.3], and "speed adjustment" is mapped to [0.4, 0.3, 0.5]) is used to generate action encoding vectors; for "parameter elements", the 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]) is used to generate parameter encoding vectors. Thus, the three types of elements of strategy 1 respectively obtain equipment encoding subsets [[0.2, 0.5, 0.1], [0.3, 0.4, 0.2]], action encoding subsets [[0.1, 0.6, 0.3], [0.4, 0.3, 0.5]], and parameter encoding subsets [[0.24, 0.1, 0.3], [0.5, 0.2, 0.4]], which together constitute the strategy encoding feature set of strategy 1; strategy 2 generates corresponding encoding subsets in the same way to form its own strategy encoding feature set.Subsequently, the server performs weighted aggregation on the encoding feature set of each strategy: according to the weight configuration of "device synergy (0.4), action responsiveness (0.3), parameter accuracy (0.3)" of the warehouse business, the device encoding subset of strategy 1 generates device aggregation features through element-wise weighted average ((0.2x0.4+0.3x0.4, 0.5x0.4+0.4x0.4, 0.1x0.4+0.2x0.4) = [0.2, 0.36, 0.12]) ; the action encoding subset generates action aggregation features through weighted average ((0.1x0.3+0.4x0.3, 0.6x0.3+0.3x0.3, 0.3x0.3+0.5x0.3) = [0.15, 0.27, 0.24]) ; the parameter encoding subset generates parameter aggregation features through weighted average ((0.24x0.3+0.5x0.3, 0.1x0.3+0.2x0.3, 0.3x0.3+0.4x0.3) = [0.222, 0.09, 0.21]) ; then, the three types of aggregation features are spliced ([0.2, 0.36, 0.12, 0.15, 0.27, 0.24, 0.222, 0.09, 0.21]) to obtain the regulation strategy features of strategy 1. Strategy 2 aggregates the encoding subsets of devices (fresh air system, air conditioner No. 3, dehumidifier No. 1), actions (turn on, keep, turn off), and parameters (26℃) according to the same weight rule to generate its own regulation strategy features. Finally, the server integrates the regulation strategy features of all regulation strategies (such as strategy 1, strategy 2) in order to form a regulation strategy feature set (such as [strategy 1 feature vector, strategy 2 feature vector,...]), which provides standardized strategy feature input for the subsequent strategy generation module's gating loop processing. Through the whole process of element deconstruction-encoding-weighted aggregation-feature integration, the server realizes the conversion of the warehouse regulation strategy from "natural language instructions" to "machine understandable high-dimensional features", laying a foundation for accurate matching of the current temperature and humidity environment.

[0063] In the embodiment of the present application, the current temperature and humidity time series data includes a plurality of temperature and humidity sampling points; the feature encoding of 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 regulation model to obtain the temperature and humidity time series feature set corresponding to the current temperature and humidity time series data and the current device operation log feature corresponding to the current device operation log can be implemented by the following example.

[0064] Based on the plurality of temperature and humidity sampling points in the current temperature and humidity time sequence data, a plurality of temperature and humidity time interval intervals are generated, and the feature extraction module in the target temperature and humidity regulation model is used to map the plurality of temperature and humidity time interval intervals to obtain the time interval initial features corresponding to the plurality of temperature and humidity time interval intervals respectively, and the plurality of time interval initial features are injected with time interval time sequence identifiers respectively to obtain a plurality of time interval enhanced features; the time interval time sequence identifier refers to the information of the position of a temperature and humidity time interval in the current temperature and humidity time sequence data; the temperature and humidity time interval is composed of one or more temperature and humidity sampling points in the plurality of temperature and humidity sampling points;

[0065] The plurality of time interval enhanced features are encoded to obtain time interval encoding features corresponding to the plurality of temperature and humidity time interval intervals respectively, and the temperature and humidity time sequence feature set corresponding to the current temperature and humidity time sequence data is generated according to the time interval encoding features corresponding to the plurality of temperature and humidity time interval intervals respectively;

[0066] The current device operation log is decomposed into a plurality of log semantic units by the feature extraction module in the target temperature and humidity regulation model;

[0067] The plurality of log semantic units are mapped to semantic vectors to obtain semantic unit vectors corresponding to the plurality of log semantic units respectively, and the plurality of semantic unit vectors are injected with log position identifiers respectively to obtain a plurality of semantic enhanced vectors; the log position identifier refers to the information of the position of a log semantic unit in the current device operation log;

[0068] The plurality of semantic enhanced vectors are encoded to obtain log semantic unit features corresponding to the plurality of log semantic units respectively, and the current device operation log features corresponding to the current device operation log are generated according to the log semantic unit features corresponding to the plurality of log semantic units respectively.

[0069] In the embodiment of the present application, in the e-commerce warehouse temperature and humidity regulation scene, the feature extraction module of the server-driven target temperature and humidity regulation model encodes the features of the current temperature and humidity time series data and the device operation log as follows: temperature and humidity time series data feature encoding (generating temperature and humidity time series feature set), the current temperature and humidity time series data is taken from the warehouse shelf sensor, collected every 5 minutes, forming a sequence containing 4 temperature and humidity sampling points: [28℃, 65%RH], [29℃, 63%RH], [30℃, 62%RH], [31℃, 61%RH]. The server generates two temperature and humidity period intervals according to the rule of "aggregating every 3 sampling points into a period": period 1: containing the first 3 sampling points (temperature rising from 28-29-30℃, humidity decreasing from 65-63-62%RH); period 2: containing the 4th sampling point (temperature 31℃, humidity 61%RH, independent segment). The feature extraction module performs feature mapping on each period: period 1: calculate the temperature mean 29℃, humidity mean 63.33%RH, humidity fluctuation variance 1.56, encode these statistics into a low-dimensional vector (such as [0.2, 0.5, 0.3]), generate period initial features; period 2: calculate temperature 31℃, humidity 61%RH (no fluctuation), encode as vector (such as [0.1, 0.4, 0.2]), generate period initial features. Inject period time sequence identifier (identify the position of the period in the overall time sequence) to the period initial features: period 1 corresponds to "the first period", concatenate position code [1, 0] after the initial features to get period enhanced features (such as [0.2, 0.5, 0.3, 1, 0]); period 2 corresponds to "the second period", concatenate position code [0, 1], get period enhanced features (such as [0.1, 0.4, 0.2, 0, 1]). Perform feature encoding on the two period enhanced features (use Transformer encoding layer to learn the time sequence association between periods): after encoding the period 1 enhanced features, the trend features of "temperature continuously rising, humidity continuously decreasing" are highlighted; after encoding the period 2 enhanced features, the abnormal information of "single point rising 31℃" is strengthened. Finally, concatenate the encoded features of the two periods in sequence (such as [trend feature vector, abnormal feature vector]), generate the temperature and humidity time series feature set corresponding to the current temperature and humidity time series data, which fully describes the time series fluctuation, trend change and period correlation of warehouse temperature and humidity. The feature encoding of the device operation log (generating the current device operation log feature) The current device operation log is the text recorded by the warehouse device management system: "air conditioner No. 1 9:00-10:00 26℃ refrigeration; dehumidifier No. 2 10:00 after medium speed".The server disassembles the log into 6 log semantic units: unit 1: "air conditioner 1" (device identification); unit 2: "9:00-10:00" (operation period); unit 3: "26℃ refrigeration" (operation parameter); unit 4: "dehumidifier 2" (device identification); unit 5: "after 10:00" (operation period); and unit 6: "medium-speed wind" (operation parameter). Semantic vector mapping (using the word vector layer of the pre-trained BERT model) is performed on each semantic unit: "air conditioner 1" is mapped to the vector [0.2, 0.5, 0.3]; "26℃ refrigeration" is mapped to the vector [0.1, 0.4, 0.2]; and the remaining units generate semantic unit vectors in the same manner. The semantic unit vectors are injected with log position identification (identifying the order of the units in the log): unit 1 (air conditioner 1) corresponds to the "first semantic unit", and the position code [1, 0, 0, 0, 0, 0] is concatenated after the vector to obtain a semantic enhanced vector (such as [0.2, 0.5, 0.3, 1, 0, 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 concatenated to generate a semantic enhanced vector; and so on, to complete the generation of enhanced vectors for all 6 units. Feature encoding (using an LSTM layer to learn the semantic association within the log) is performed on the semantic enhanced vectors: after the enhanced vectors of unit 1 (air conditioner 1) and unit 3 (26℃ refrigeration) are encoded by LSTM, the device-parameter association of "air conditioner 1 running at 26℃ refrigeration" is strengthened; after the enhanced vectors of unit 4 (dehumidifier 2) and unit 6 (medium-speed wind) are encoded, the device-parameter association of "dehumidifier 2 running at medium-speed wind" is strengthened; and after the enhanced vectors of unit 2 (9:00-10:00) and unit 5 (after 10:00) are encoded, the time period association of "air conditioner 1 running in the morning and dehumidifier 2 running after noon" is strengthened. Finally, the encoded features of the 6 log semantic units are aggregated in sequence (such as weighted average or concatenation) to generate the current device operation log feature corresponding to the current device operation log, which fully describes the core information of the device, such as start-stop state, operation parameter, and time association. Through the "period division-feature mapping-time sequence encoding" and "log disassembly-semantic mapping-semantic encoding" double processes of the feature extraction module, the server realizes the precise digital expression of the warehouse temperature and humidity dynamics and the device operation state, providing fine-grained and multi-dimensional feature input for the subsequent adaptive analysis of the strategy generation module.

[0070] In the embodiments of the present application, the current temperature and humidity time sequence data includes a plurality of temperature and humidity sampling points, and the number of target control strategies is a plurality, and the target control strategy includes a first control strategy; the method further provides the following implementation.

[0071] generate a plurality of temperature and humidity time interval based on a plurality of temperature and humidity sampling points in the current temperature and humidity time sequence data, obtain a temperature and humidity time interval feature set corresponding to the plurality of temperature and humidity time interval and a first strategy reference feature corresponding to the first regulation strategy; the temperature and humidity time interval is composed of one or more temperature and humidity sampling points in the plurality of temperature and humidity sampling points;

[0072] perform gating cycle processing on the temperature and humidity time interval feature set and the first strategy reference feature respectively through the strategy generation module, obtain the temperature and humidity time interval gating cycle data set corresponding to the first regulation strategy; each gating cycle value in the temperature and humidity time interval gating cycle data set is used to represent the adaptation relationship between the first regulation strategy and a temperature and humidity time interval;

[0073] output the temperature and humidity time interval adaptation probability value corresponding to each temperature and humidity time interval according to the temperature and humidity time interval gating cycle data set, obtain the target temperature and humidity time interval associated with the first regulation strategy in the plurality of temperature and humidity time intervals according to the temperature and humidity time interval adaptation probability value, and set the first regulation strategy for the target temperature and humidity time interval.

[0074] In the embodiment of the present application, in the e-commerce warehouse temperature and humidity fine control scene, the server executes the time period level adaptation process of the multi-target control strategy. Taking the first control strategy ("air conditioner No. 1 is adjusted to 24 DEG C refrigeration, and the dehumidifier No. 2 is switched to medium speed") and the current temperature and humidity time series data as an example, the following steps are carried out: step 1: generating temperature and humidity time period intervals and obtaining a feature set, the current temperature and humidity time series data is taken from the shelf area sensor, collected every 5 minutes, forming a sequence containing 4 temperature and humidity sampling points: [28 DEG C, 65% RH] (9:00), [29 DEG C, 63% RH] (9:05), [30 DEG C, 62% RH] (9:10), [31 DEG C, 61% RH] (9:15). The server generates two temperature and humidity time period intervals according to the rule of "aggregating every 3 continuous sampling points into a time period": time period 1: covering 3 sampling points of 9:00-9:10 (temperature continuously rising from 28-29-30 DEG C, humidity continuously decreasing from 65-63-62% RH); time period 2: covering 1 sampling point of 9:15 (temperature suddenly rising to 31 DEG C, humidity 61% RH). Through the feature extraction module of the target temperature and humidity control model, the "feature mapping-time sequence identification injection-encoding" process is performed on the two time periods: for time period 1, the temperature average 29 DEG C, humidity average 63.33% RH and other statistics are calculated and encoded into a vector, and the enhanced features are obtained by injecting the time sequence identification of "the first time period"; for time period 2, the abnormal information of "31 DEG C single point sudden rise" is extracted and encoded, and the "second time period" identification is injected. Finally, the temperature and humidity time period interval feature set is generated, wherein the time period 1 feature focuses on the "temperature and humidity continuous fluctuation trend", and the time period 2 feature focuses on the "single point anomaly". At the same time, the server retrieves the first control strategy reference feature from the control strategy library: through the "strategy element deconstruction (decomposed into 'air conditioner No. 1', '24 DEG C refrigeration', 'dehumidifier No. 2','medium speed' and other elements)-encoding (device, action, parameter embedded into vector)-weighted aggregation (integrated according to device cooperation, action response, parameter precision weight)" process, the first strategy reference feature vector containing control logic is generated. Step 2: the gating loop processing of the strategy generation module, the server drives the strategy generation module to input the temperature and humidity time period interval feature set (time period 1, time period 2 feature) and the first strategy reference feature one by one, and performs gating loop processing: for the time period 1 feature and the first strategy reference feature: the gating loop operator of the strategy generation module takes the "trend feature vector" of time period 1 as the environment sensing vector (describing the temperature and humidity fluctuation law of 9:00-9:10), and takes the first strategy reference feature as the strategy index vector (describing the control logic of "24 DEG C refrigeration + medium speed dehumidification"). Through the "vector multiplication (quantitative strategy and environment preliminary association)-inverse weight (balance control direction weight)-nonlinear mapping (strengthen core association)-residual integration (preserve original information)" process, the gating loop value of time period 1 and the first strategy (such as 0.8) is calculated, which quantitatively represents the adaptation strength of the two.For the period 2 characteristics and the first strategy benchmark characteristics: for the same reason, the "abnormal feature vector" of period 2 is taken as the environmental sensing vector, and the first strategy benchmark characteristics are taken as the strategy index vector; after the gated recurrent processing, the gated recurrent value of period 2 and the first strategy (such as 0.3) is obtained. Finally, the two gated recurrent values constitute the temperature and humidity period interval gated recurrent data set corresponding to the first regulation strategy, wherein each value accurately depicts the adaptation relationship between the first strategy and the corresponding period (period 1 has a higher adaptation degree). Step 3: Adaptation probability calculation and period-level strategy deployment, the server calls the decision output unit to perform Softmax nonlinear transformation on the gated recurrent data set to generate temperature and humidity period interval adaptation probability values: period 1 probability: e 0.8 ÷(e 0.8 +e 0.3 )≈0.75; period 2 probability: e 0.3 ÷(e 0.8 +e 0.3 )≈0.25. The server presets a probability threshold (such as 0.5) to filter out period 1 (probability 0.75> 0.5) as the target temperature and humidity period interval associated with the first regulation strategy. Then, the server binds the execution instructions of the first regulation strategy to the time interval corresponding to period 1 (9:00-9:10): through the warehouse Internet of Things platform, send the "start cooling mode during 9:00-9:10, set temperature to 24℃" instruction to air conditioner No. 1, and send the "adjust the air speed to medium speed during 9:00-9:10" instruction to dehumidifier No. 2, to realize the precise binding and execution of "strategy-period-equipment". Through the whole process of "period-level feature disassembly-strategy-period gated adaptation-probability screening-period deployment", the server realizes the refinement of the warehouse temperature and humidity regulation from "global strategy matching" to "period-level precise intervention". For example, in the period of 9:00-9:10 when the temperature and humidity continue to deteriorate, the first regulation strategy is preferentially deployed to quickly suppress temperature rise and humidity decrease; while the 9:15 single-point sudden rise period triggers other strategies (such as fresh air system linkage), maximizing the regulation efficiency and energy efficiency ratio, and ensuring the stability of the goods storage environment.

[0075] In the embodiment of the application, the target temperature and humidity regulation model is obtained by the following method, which can be implemented by the following example.

[0076] The sample temperature and humidity time series data and the device operation log associated with the sample temperature and humidity time series data are input into the original temperature and humidity regulation model; the original temperature and humidity regulation model comprises an original feature extraction module and an original strategy generation module;

[0077] The original feature extraction module in the original temperature and humidity regulation model is used for feature coding 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 equipment 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 equipment log feature, to obtain a sample fusion environment feature set;

[0078] A plurality of regulation strategy instance features corresponding to regulation strategy instances are obtained, the original strategy generation module in the original temperature and humidity regulation model is used for gated recurrent processing on the sample fusion environment feature set and the regulation strategy instance feature set, to obtain a sample gated recurrent data set corresponding to the regulation strategy instance feature set; each gated recurrent value in the sample gated recurrent data set is used for representing a matching degree between the sample temperature and humidity time series data and one regulation strategy instance;

[0079] The original feature extraction module and the original strategy generation module in the original temperature and humidity regulation model are model optimized according to the sample gated recurrent data set, to obtain a target temperature and humidity regulation model, which is used for regulation strategy matching of warehouse temperature and humidity.

[0080] In the embodiment of the present application, in the e-commerce warehouse temperature and humidity regulation model training scene, the server executes the training process of the target temperature and humidity regulation model as follows: first, input the sample data to the original model: the server retrieves the sample temperature and humidity time series data of a certain week in summer (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℃, 65%RH)—11:00 (30℃, 63%RH)—13:00 (32℃, 60%RH)—15:00 (31℃, 62%RH)—17:00 (30℃, 64%RH)”) and the device operation log associated with the period (such as “air conditioner No. 1 runs at 26℃ cooling from 9:00 to 12:00; dehumidifier No. 2 adjusts the air speed to medium after 10:00; 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 regulation model (with an original feature extraction module and an original strategy generation module built-in). Second, the original feature extraction module generates a sample fusion environment feature set: for the sample temperature and humidity time series data, the server generates 3 time periods (9:00-11:00, 11:00-13:00, 13:00-15:00, and the remaining 15:00-17:00 as the 4th time period) according to the rule of “dividing a temperature and humidity time period interval every 2 hours”. The original feature extraction module performs “feature mapping-time sequence identification injection-encoding” on each time period: calculates the temperature mean of period 1 (9:00-11:00) as 29℃, the humidity mean as 64%RH, and the humidity fluctuation variance as 1, encodes and injects the “1st time period” identification to generate the period enhanced feature; similarly process the remaining 3 time periods, and finally encode to obtain the historical temperature and humidity time series feature set (characterizing the trend, fluctuation and time period correlation of temperature and humidity over time). For the device operation log, the original feature extraction module decomposes the log into 9 log semantic units such as “air conditioner No. 1”, “9:00-12:00”, “26℃ cooling”, “dehumidifier No. 2”, “after 10:00”, “air speed medium”, “fresh air system”, “14:00-16:00”, and “on”; maps them to semantic unit vectors through a pre-trained word vector model, injects the “1st semantic unit (air conditioner No. 1)” and “2nd semantic unit (9:00-12:00)” position identifications, and generates semantic enhanced vectors; aggregates the sample device log features (characterizing the device start-stop, parameters, and time correlation) after LST M encoding. The server performs “element-wise weighted splicing” on the historical temperature and humidity time series features and the sample device log features to generate a sample fusion environment feature set (carrying high-dimensional features of temperature and humidity dynamics and device real-time status).Next, the original strategy generation module generates a sample gating cycle dataset: the server retrieves historical validation effective control strategy instances (such as strategy instance A: "Air conditioner No. 1 is adjusted to 25 DEG C refrigeration, dehumidifier No. 2 is switched to medium speed"; strategy instance B: "Turn on fresh air system, keep air conditioner No. 3 running at 27 DEG C, and turn off dehumidifier No. 1"). For each strategy instance, perform the "element deconstruction (device, action, parameter) - encoding (embedding layer mapping) - weighted aggregation" process to generate a control strategy 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 gating cycle branch: the first branch takes the control strategy instance feature set as the cycle precursor, combines the sample fusion environment feature set to calculate the original gating cycle feature (quantitative strategy and environment preliminary matching), and outputs the cycle response feature after standardization and residual connection; the subsequent branch inherits the previous feature iteration processing, and finally outputs the sample gating cycle dataset (such as strategy instance A corresponding matching degree 0.85, strategy instance B corresponding 0.62, the adaptation strength of the sample temperature and humidity time sequence and each strategy instance). Finally, the target model is obtained by model optimization: the server takes the "difference between the sample gating cycle dataset and the actual optimal strategy label" as the loss function (such as cross-entropy loss, the label is the actual effective strategy instance under the sample data annotated by artificial labeling), and iteratively adjusts the parameters of the original feature extraction module and the original strategy generation module (such as optimizing the weight of feature encoding and the coefficient of gating cycle operator) through the back propagation algorithm. When the loss function converges (such as the loss decreases by less than 0.001 for 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 and control strategy in the warehouse, providing core algorithm support for intelligent temperature and humidity control in actual scenarios. Through the whole process of "sample data input - feature encoding integration - strategy instance adaptation quantification - parameter iteration optimization", the server completes the training of the target temperature and humidity control model, ensuring that the model can efficiently identify the adaptation relationship between temperature and humidity trends, device status and control strategies in actual warehouse scenarios, and achieve intelligent and accurate control.

[0081] In the embodiment of the application, the original feature extraction module includes a temperature and humidity time sequence feature extraction unit and a log description feature extraction unit. The original temperature and humidity control model is optimized according to the sample gating cycle dataset to obtain a target temperature and humidity control model, which can be implemented by the following examples.

[0082] Obtain the reference temperature and humidity time sequence feature corresponding to the sample temperature and humidity time sequence data, and optimize the temperature and humidity time sequence feature extraction unit according to the reference temperature and humidity time sequence feature and the historical temperature and humidity time sequence feature set;

[0083] obtaining the corresponding reference device log feature of the device operation log, and performing model optimization on the log description feature extraction unit according to the reference device log feature and the sample device log feature;

[0084] obtaining the corresponding standard control strategy template of the sample temperature and humidity time sequence data, and generating a strategy matching error value according to the standard control strategy template and the sample gate cycle data set;

[0085] performing model optimization on the original strategy generation module in the original temperature and humidity control model according to the strategy matching error value, to obtain a target temperature and humidity control model.

[0086] In the embodiment of the application, for example, in the optimization stage of the e-commerce warehouse target temperature and humidity control model, the server iteratively optimizes the original feature extraction module (including the temperature and humidity time sequence feature extraction unit and the log description feature extraction unit) and the original strategy generation module as follows: 1. Optimization of the temperature and humidity time sequence feature extraction unit: the server retrieves sample temperature and humidity time sequence data (such as the temperature and humidity sequence collected every hour from 9:00 to 17:00 in the summer: [28℃, 65%RH] (9:00), [29℃, 63%RH] (10:00), [30℃, 62%RH] (11:00), [31℃, 61%RH] (12:00), [30℃, 63%RH] (13:00)) from the warehouse historical database, and obtains the corresponding reference temperature and humidity time sequence feature (an ideal feature vector manually annotated by a domain expert in combination with temperature and humidity fluctuation rules and goods storage threshold values, such as [0.8, 0.2, 0.9] representing “temperature rise trend intensity”, “humidity drop trend intensity” and “period abnormality degree” respectively). The temperature and humidity time sequence feature extraction unit of the original temperature and humidity control model performs “period division (aggregating every 2 hours into period 1: 9-11, period 2: 11-13)—feature mapping (calculating the temperature mean value 29℃ and humidity mean value 64%RH of period 1 and encoding as a vector)—time sequence identification injection (adding “first period” position coding to the period 1 feature)—Transformer encoding” on the sample data, to generate a historical temperature and humidity time sequence feature set (such as period 1 feature [0.7, 0.3, 0.8] and period 2 feature [0.9, 0.1, 0.7]). The server calculates the mean square error (MSE) of the historical temperature and humidity time sequence feature set and the reference temperature and humidity time sequence feature (such as the MSE of period 1 feature and the reference is (0.8-0.7) 2 +(0.2-0.3) 2 +(0.9-0.8) 2= 0.03), the neural network weights of the temperature and humidity time series feature extraction unit are adjusted by the back propagation algorithm (such as optimizing the time window parameters of the period aggregation rule, the weight matrix of the feature mapping embedding layer), the feature coding error is minimized, and the model optimization of the unit is completed. 2. The optimization of the log description feature extraction unit is described. The server calls the device operation log (text record: "Air conditioner No. 1 runs at 26°C cooling from 9:00 to 11:00; dehumidifier No. 2 adjusts the wind speed to medium after 10:00") associated with the sample temperature and humidity time series data, and obtains the reference device log feature (the core information vector labeled by the device engineer, such as [0.9, 0.1, 0.8] representing "air conditioner No. 1 cooling effectiveness" "dehumidifier No. 2 wind speed adaptability" "device period coordination degree" respectively). The log description feature extraction unit of the original feature extraction module performs "semantic unit decomposition (decomposed into 6 semantic units of 'air conditioner No. 1' '9:00-11:00' '26°C cooling' 'dehumidifier No. 2' 'after 10:00' 'wind speed medium') - semantic vector mapping (generate word vectors for each unit by pre-training BERT model, such as 'air conditioner No. 1' mapped to [0.2, 0.5, 0.3]) - position identification injection (add "the first semantic unit" position code [1, 0, 0, 0, 0, 0] to the 'air conditioner No. 1' unit) - LSTM encoding (learning the semantic association within the log)", to generate the sample device log feature (such as the aggregated backward vector [0.8, 0.2, 0.7]). The server calculates the cosine similarity error between the sample device log feature and the reference device log feature (such as similarity 0.8x0.9+0.2x0.1+0.7x0.8=1.3, normalized error ), the word vector embedding layer weight of the log description feature extraction unit and the loop kernel parameter of the LSTM layer are adjusted through back propagation to optimize the coding accuracy of the log semantics and the core operation and maintenance information, and the unit optimization is completed. 3. Optimization of the original strategy generation module, the server obtains the standard control strategy template corresponding to the sample temperature and humidity time series data (the actual optimal strategy labeled by the warehouse operation team, such as "air conditioner No. 1 is adjusted to 25 DEG C refrigeration, and dehumidifier No. 2 is kept at medium speed") and generates the strategy benchmark features of the template (the vector after "element deconstruction-coding-weighted aggregation" is [0.9, 0.1, 0.8]) through the feature extraction module. The original strategy generation module executes the 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 (containing the features of the standard strategy template): the first branch takes 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 branch iteratively processes, and finally generates a 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 strategy template, and p i is the matching degree 0.7 predicted by the model), and the loss value L = -1 x log 0.7 is obtained. 0.357. The gated loop branch weight of the original strategy generation module and the activation function parameter of the fully connected layer are adjusted through back propagation to minimize the strategy matching error, and the module optimization is completed. After the above three-stage iterative optimization, the server updates all trainable parameters of the original temperature and humidity control model to obtain the target temperature and humidity control model. The model has high precision in temperature and humidity trend coding, log semantics analysis, and strategy adaptation quantization, can quickly match the optimal control strategy in the actual warehouse scene, and ensures the stability of the goods storage environment and the optimal energy efficiency of the equipment.

[0087] In the embodiment of the present application, the generation of the strategy matching error value according to the standard control strategy template and the sample gated loop data set can be implemented by the following example.

[0088] According to the sample gated loop data set, output the sample adaptation probability value corresponding to each of the plurality of control strategy instances;

[0089] Respectively, the plurality of sample adaptation probability values are exponentially inverse domain converted to obtain a plurality of matching degree inverse domain quantization values;

[0090] determine a plurality of policy matching identification values corresponding to the plurality of regulation policy instances respectively based on the standard regulation policy template and the sample adaptation probability value; the policy matching identification value comprises a matching identification value or a non-matching identification value, the matching identification value is used to represent that the standard regulation policy template contains a policy template matched with the regulation policy instance corresponding to the matching identification value, and the non-matching identification value is used to represent that the standard regulation policy template does not contain a policy template matched with the regulation policy instance corresponding to the non-matching identification value;

[0091] generate a policy matching error value according to the policy matching identification values corresponding to the plurality of regulation policy instances respectively and the plurality of matching degree inverse domain quantization values.

[0092] In the embodiment of the present application, for example, in the e-commerce warehouse model optimization scenario, when the server processes the policy matching error value generation process, taking the sample temperature and humidity time sequence data (shelf area 9:00-11:00 temperature and humidity sequence [28-29-30℃, 65-63-62% RH]), the standard regulation policy template ("air conditioner No. 1 is adjusted to 25℃ for refrigeration, and dehumidifier No. 2 is kept at medium speed") and the regulation policy instances A (same as the standard template), B ("turn on the fresh air system + air conditioner No. 3 keep running at 27℃") and C ("turn off air conditioner No. 1 + turn on dehumidifier No. 1 at high speed") as examples, the following steps are executed: Step 1: output the sample adaptation probability value, the server calls the decision output unit of the original policy generation module to perform Softmax nonlinear transformation on the sample gated recurrent data set (containing the adaptation quantization values [1.6, 1.2, 0.8] of strategies A, B and C) obtained through the gated recurrent processing, and calculate 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; finally generate the sample adaptation probability value set {A: 0.45, B: 0.32, C: 0.21}, and the quantization model calculates the matching confidence of each strategy and the sample temperature and humidity environment. Step 2: exponential inverse domain conversion to obtain the matching degree inverse domain quantization value, the server performs exponential inverse domain conversion (the conversion formula is on each sample adaptation probability value (the error sensitivity of low probability strategies is enhanced 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; obtain the matching degree inverse domain quantization value set {A:0.637,B:0.726,C:0.810}, so that the error contribution of low probability strategy is more easily captured in subsequent loss calculation. Step 3: Determine the strategy matching identifier value. The server compares the semantic consistency of the standard control strategy template ("Air conditioner No. 1 adjusts to 25℃ cooling + dehumidifier No. 2 fan speed medium") with each control strategy instance: Strategy A is completely consistent with the standard template equipment, action, and parameters - marked as matching identifier value (1); the equipment combination and action instructions of strategy B (fresh air + air conditioner No. 3) and strategy C (turn off air conditioner No. 1 + dehumidifier No. 1 high speed) do not overlap with the standard template - both are marked as unmatched identifier value (0); generate the strategy matching identifier value set {A:1,B:0,C:0}, and clarify the supervision label of "whether the standard template and strategy instance match". Step 4: Generate the strategy matching error value. The server uses the weighted cross-entropy loss calculation model to predict the error of the supervision label. The formula is: , where w i As the importance weights of the strategies (since strategy A matches the standard template, its weight is set to 2; B and C are set to 1), y i For the strategy matching identifier value, q i This is the inverse domain quantization value of the matching degree. Substituting the values ​​into the calculation: 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 value is minimized through backpropagation during training to guide the model parameters to iterate towards the "precise matching standard strategy"). Through a hierarchical process of "probability output - inverse domain transformation - labeling - loss calculation", the server accurately quantifies the prediction error of the model on the "matching relationship between standard policy template and regulation instance", provides gradient direction for the parameter iteration of the original policy generation module, ensures that the model learns accurate correlation rules in the policy adaptation task, and finally outputs a highly generalizable target temperature and humidity regulation model.

[0093] This 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 deep learning-based intelligent temperature and humidity control method for warehouses. Figure 2 As shown, Figure 2A structural block diagram of the computer device 100 is provided for the embodiments of the present application. The computer device 100 comprises a memory 111, a processor 112 and a communication unit 113. For realizing the transmission or interaction of data, the memory 111, the processor 112 and the communication unit 113 are electrically connected with each other directly or indirectly. For example, the electrical connection between these elements can be realized by one or more communication buses or signal lines.

[0094] The foregoing description is made with reference to specific embodiments for purposes of illustration only. The illustrative discussion above, however, is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Rather, various modifications and changes can be effected therein by those skilled in the art. These embodiments were chosen and described by way of example for the best mode contemplated of carrying out the disclosure, and the full scope of the disclosure is not to be limited by any aspect of these described embodiments.

Claims

1. A warehouse temperature and humidity intelligent regulation method based on deep learning, characterized in that, The method comprises the following steps: inputting current temperature and humidity time sequence data and current equipment operation log associated with the current temperature and humidity time sequence data into a target temperature and humidity control model; the target temperature and humidity control model comprises a feature extraction module and a strategy generation module; the feature extraction module in the target temperature and humidity control model is used for feature coding of the current temperature and humidity time sequence data and the current equipment operation log, to obtain temperature and humidity time sequence features corresponding to the current temperature and humidity time sequence data, current equipment operation log features corresponding to the current equipment operation log, and performing feature integration operation on the temperature and humidity time sequence features and the current equipment operation log features to obtain a fusion environment feature set; obtain a plurality of control strategy features corresponding to a plurality of control strategies, and perform gated recurrent processing on the fusion environment feature set and the control strategy feature set through the strategy generation module to obtain a gated recurrent data set corresponding to the control strategy feature set; each gated recurrent value in the gated recurrent data set is used to represent the adaptive relationship between the current temperature and humidity time sequence data and one control strategy; output the adaptive probability values corresponding to the plurality of control strategies respectively according to the gated recurrent data set, and set a target control strategy in the plurality of control strategies for the current temperature and humidity time sequence data according to the adaptive probability values; the current temperature and humidity time sequence data comprises a plurality of temperature and humidity sampling points; the feature extraction module in the target temperature and humidity control model is used for feature coding of the current temperature and humidity time sequence data and the current equipment operation log, to obtain temperature and humidity time sequence features corresponding to the current temperature and humidity time sequence data, current equipment operation log features corresponding to the current equipment operation log, and performing feature integration operation on the temperature and humidity time sequence features and the current equipment operation log features to obtain a fusion environment feature set; based on the plurality of temperature and humidity sampling points in the current temperature and humidity time sequence data, a plurality of temperature and humidity time interval intervals are generated, and the feature extraction module in the target temperature and humidity control model is used for feature mapping of the plurality of temperature and humidity time interval intervals to obtain time interval initial features corresponding to the plurality of temperature and humidity time interval intervals respectively, and injecting time interval time sequence identifiers into a plurality of time interval initial features respectively to obtain a plurality of time interval enhanced features; the time interval time sequence identifier refers to the information of the position of a temperature and humidity time interval in the current temperature and humidity time sequence data; the temperature and humidity time interval is composed of one or more temperature and humidity sampling points in the plurality of temperature and humidity sampling points; feature coding is performed on the plurality of time interval enhanced features to obtain time interval coding features corresponding to the plurality of temperature and humidity time interval intervals respectively, and the temperature and humidity time sequence feature set corresponding to the current temperature and humidity time sequence data is generated according to the time interval coding features corresponding to the plurality of temperature and humidity time interval intervals respectively; the current equipment operation log is decomposed into a plurality of log semantic units through the feature extraction module in the target temperature and humidity control model; semantic vector mapping is performed on the plurality of log semantic units to obtain semantic unit vectors corresponding to the plurality of log semantic units respectively, and log position identifiers are injected into a plurality of semantic unit vectors respectively to obtain a plurality of semantic enhanced vectors; the log position identifier refers to the information of the position of a log semantic unit in the current equipment operation log; The plurality of semantic enhancement vectors are feature encoded to obtain log semantic unit features corresponding to the plurality of log semantic units respectively, and the current device running log features corresponding to the current device running log are generated according to the log semantic unit features corresponding to the plurality of log semantic units respectively.

2. The method of claim 1, wherein, The fusion environment feature set and the regulation strategy feature set are processed by the strategy generation module to obtain a gating cycle data set corresponding to the regulation strategy feature set, including: In the strategy generation module, the fusion environment feature set and the regulation strategy feature set are processed by the gating cycle to obtain original gating cycle features; The original gating cycle features are standardized to obtain gating cycle standardized features; The gating cycle standardized features are processed by a full connection layer and an activation function to obtain gating cycle conduction features, a residual connection operation is performed on the gating cycle standardized features and the gating cycle conduction features to obtain gating cycle coupling features, and the gating cycle coupling features are standardized to obtain the gating cycle data set corresponding to the regulation strategy feature set.

3. The method of claim 1, wherein, The strategy generation module includes a plurality of gating cycle branches, the plurality of gating cycle branches are cascaded, and the plurality of gating cycle branches include a target gating cycle branch; the fusion environment feature set and the regulation strategy feature set are processed by the strategy generation module to obtain a gating cycle data set corresponding to the regulation strategy feature set, including: The cycle predecessor features corresponding to the target gating cycle branch and the fusion environment feature set are taken as cycle input features of the target gating cycle branch, and the cycle input features of the target gating cycle branch are processed by the gating cycle in the target gating cycle branch to obtain cycle response features of the target gating cycle branch; if the target gating cycle branch is the first gating cycle branch in the plurality of gating cycle branches, the cycle predecessor features corresponding to the target gating cycle branch are the regulation strategy feature set; The cycle response features of the target gating cycle branch are taken as cycle predecessor features corresponding to a subsequent gating cycle branch, the cycle predecessor features corresponding to the subsequent gating cycle branch and the fusion environment feature set are taken as cycle input features of the subsequent gating cycle branch, and the cycle input features of the subsequent gating cycle branch are processed by the gating cycle in the subsequent gating cycle branch to obtain cycle response features of the subsequent gating cycle branch, until cycle response features of a last gating cycle branch in the plurality of gating cycle branches are obtained, the cycle response features of the last gating cycle branch are determined as the gating cycle data set corresponding to the regulation strategy feature set; the subsequent gating cycle branch is a gating cycle branch subsequent to the target gating cycle branch.

4. The method of claim 3, wherein, The cycle predecessor feature corresponding to the target gated recurrent branch and the set of fusion environment features are taken as cycle input features of the target gated recurrent branch, the cycle input features of the target gated recurrent branch are processed in the target gated recurrent branch, and a cycle response feature of the target gated recurrent branch is obtained, including: In the target gated recurrent branch, the cycle predecessor feature is taken as a policy index vector in a gated recurrent operator, the set of fusion environment features is taken as an environment sensing vector in the gated recurrent operator, and the set of fusion environment features is taken as a device state vector in the gated recurrent operator; A multiplication result of the policy index vector and the environment sensing vector is determined as a first interaction feature through the gated recurrent operator; A multiplication result of the first interaction feature and a multiplicative inverse of the number of regulation directions is determined as a second interaction feature; The second interaction feature is converted into a first mapping feature based on a nonlinear transformation subfunction in the gated recurrent operator; 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 recurrent feature; A neuron activation amount in the original gated recurrent feature is obtained, and an average value and a standard deviation of the neuron activation amount are determined; A difference value between the original gated recurrent feature and the average value is determined as a centralization feature amount; A standard deviation and a preset standardization parameter are added, and a square root of the addition result is obtained as a standardization coefficient; A multiplication result between the centralization feature amount and a multiplicative inverse of the standardization coefficient is determined as a standardized feature; The standardized feature is processed through a fully connected layer and an activation function to obtain a logical conduction feature, a residual connection operation is performed on the standardized feature and the logical conduction feature to obtain a gated recurrent environment coupling feature, and the gated recurrent environment coupling feature is standardized to obtain the cycle response feature of the target gated recurrent branch.

5. The method of claim 1, wherein, The target temperature and humidity regulation model further includes a decision output unit, the adaptive probability value corresponding to each of the plurality of regulation strategies is output according to the gated recurrent data set, and the target regulation strategy in the plurality of regulation strategies is set for the current temperature and humidity time series data according to the adaptive probability value, including: The gated recurrent data set is subjected to confidence standardization processing based on a nonlinear transformation function of the decision output unit to obtain the adaptive probability value corresponding to each of the plurality of regulation strategies; An adaptive probability value exceeding a preset probability threshold in the plurality of adaptive probability values is determined as a target adaptive probability value, and the regulation strategy corresponding to the target adaptive probability value is deployed with the current temperature and humidity time series data.

6. The method of claim 1, wherein, The regulation strategy feature set corresponding to the plurality of regulation strategies is obtained, including: A plurality of regulation strategies are obtained, and the plurality of regulation strategies are respectively deconstructed to obtain strategy elements corresponding to the plurality of regulation strategies, respectively. The feature extraction module in the target temperature and humidity regulation model encodes multiple strategy elements to obtain a strategy encoding feature set corresponding to each of the multiple strategy elements; The multiple strategy encoding feature sets are respectively weighted and aggregated to obtain regulation strategy features corresponding to multiple regulation strategies; The multiple regulation strategy features are integrated to obtain regulation strategy feature sets corresponding to the multiple regulation strategies.

7. The method of claim 1, wherein, The current temperature and humidity time series data includes multiple temperature and humidity sampling points, and the target regulation strategy includes a first regulation strategy; the method further comprises: Based on the multiple temperature and humidity sampling points in the current temperature and humidity time series data, multiple temperature and humidity time interval ranges are generated, and a temperature and humidity time interval range feature set corresponding to the multiple temperature and humidity time interval ranges and a first strategy reference feature corresponding to the first regulation strategy are obtained; the temperature and humidity time interval range is composed of one or more temperature and humidity sampling points in the multiple temperature and humidity sampling points; The strategy generation module performs gated recurrent processing on the temperature and humidity time interval range feature set and the first strategy reference feature to obtain a temperature and humidity time interval range gated recurrent data set corresponding to the first regulation strategy; each gated recurrent value in the temperature and humidity time interval range gated recurrent data set is used to represent the adaptation relationship between the first regulation strategy and a temperature and humidity time interval range; According to the temperature and humidity time interval range gated recurrent data set, temperature and humidity time interval range adaptation probability values corresponding to the multiple temperature and humidity time interval ranges are output, and a target temperature and humidity time interval range associated with the first regulation strategy is obtained from the multiple temperature and humidity time interval ranges according to the temperature and humidity time interval range adaptation probability values, and the first regulation strategy is set for the target temperature and humidity time interval range.

8. The method of claim 1, wherein, The target temperature and humidity regulation model is obtained by the following method, comprising: Sample temperature and humidity time series data and device operation logs associated with the sample temperature and humidity time series data are input into an original temperature and humidity regulation model; the original temperature and humidity regulation model includes an original feature extraction module and an original strategy generation module; The original feature extraction module in the original temperature and humidity regulation model encodes the sample temperature and humidity time series data and the device operation logs 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 device operation logs, and performs feature integration on the historical temperature and humidity time series feature set and the sample device log feature to obtain a sample fusion environment feature set; A regulation strategy instance feature set corresponding to multiple regulation strategy instances is obtained, and the original strategy generation module in the original temperature and humidity regulation model performs gated recurrent processing on the sample fusion environment feature set and the regulation strategy instance feature set to obtain a sample gated recurrent data set corresponding to the regulation strategy instance feature set; each gated recurrent value in the sample gated recurrent data set is used to represent the matching degree between the sample temperature and humidity time series data and a regulation strategy instance; According to the sample gating cycle dataset, the original feature extraction module and the original strategy generation module in the original temperature and humidity regulation model are model optimized to obtain a target temperature and humidity regulation model, and the target temperature and humidity regulation model is used for warehouse temperature and humidity regulation strategy matching.

9. A server system, characterized by The system comprises a server configured to perform the method of any one of claims 1-8.

Images