Exhibition hall multi-device linkage energy-saving scheduling method and system based on reinforcement learning

Abstract

The application provides a showroom multi-device linkage energy-saving scheduling method and system based on reinforcement learning, and relates to the technical field of showroom energy-saving scheduling. The method comprises the following steps: obtaining a first micro-environment parameter located in an independent micro-environment showcase, a second wall heat flux of a phase change heat storage wall surface, and a third global environment parameter below a radiation refrigeration ceiling, to obtain an initial environment parameter set; performing normalization processing on each parameter in the initial environment parameter set respectively, mapping the normalized environment parameter set to a feature manifold space, calculating the geodesic distance of the parameter vector corresponding to each sampling time in the normalized environment parameter set in the feature manifold space, and constructing a space distance matrix; performing multidimensional scaling transformation on the space distance matrix, extracting low-dimensional geometric coordinates that maintain the global space topological structure, and obtaining an environment state embedding vector. The application improves the accuracy and stability of the multi-device linkage scheduling of the showroom.

Description

Technical Field

[0001] This invention relates to the field of energy-saving scheduling technology for exhibition halls, and in particular to a method and system for multi-device linkage energy-saving scheduling in exhibition halls based on reinforcement learning. Background Technology

[0002] In the constant temperature and humidity exhibition halls for geological specimens in high-altitude and cold regions, various environmental control devices are usually arranged, such as independent microenvironment display cases, phase change heat storage walls, radiant cooling ceilings, and semiconductor temperature control bases. Existing exhibition hall environmental control methods often focus on independent local or global control, relying mainly on global sensors for conventional feedback control, with relatively limited consideration for the interaction between the microenvironment and the macroenvironment.

[0003] This approach has certain limitations in practical applications: it is difficult to fully quantify the complex thermo-humidity coupling and permeation effects between the display case microenvironment, the phase change heat storage wall, and the radiant cooling ceiling. Due to the lack of effective quantification of the spatial thermo-humidity permeation topology and cumulative conduction resistance, intelligent scheduling algorithms such as reinforcement learning struggle to construct accurate global spatial feature representations when acquiring state inputs, thus failing to accurately reconstruct the dynamic transmission path of thermo-humidity energy in the three-dimensional space of the exhibition hall. This leads to fluctuations in the action strategies of multiple devices working together, and the semiconductor temperature control base and radiant cooling ceiling may over-adjust; this not only easily causes cold energy redundancy and increased operating energy consumption, but under certain operating conditions, it may also cause condensation risks on the inner walls of the display case glass due to local temperature and humidity fluctuations, which is detrimental to the long-term safe preservation of precious specimens.

[0004] This invention aims to accurately quantify the spatial heat and humidity conduction network by introducing characteristic manifold spatial mapping and multi-dimensional scale transformation, and by combining time series smoothing processing, effectively improves the above-mentioned control limitations, and achieves stable linkage of multiple devices and improved energy efficiency. Summary of the Invention

[0005] This invention provides a reinforcement learning-based method and system for energy-saving scheduling of multiple devices in an exhibition hall, which improves the accuracy and stability of multi-device coordinated scheduling in an exhibition hall.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a reinforcement learning-based multi-device linkage energy-saving scheduling method for exhibition halls, the method comprising: Step 1: Obtain the first microenvironmental parameters located inside the independent microenvironment display case, the second wall heat flux on the surface of the phase change heat storage wall, and the third global environmental parameters below the radiant cooling ceiling to obtain the initial environmental parameter set; Step 2: Normalize each parameter in the initial environmental parameter set, map the normalized environmental parameter set to the feature manifold space, calculate the geodesic distance of the parameter vector corresponding to each sampling time in the normalized environmental parameter set in the feature manifold space, and construct the spatial distance matrix. Step 3: Perform multi-dimensional scaling transformation on the spatial distance matrix to extract low-dimensional geometric coordinates that preserve the global spatial topology, and obtain the environment state embedding vector; Step 4: Embed the environmental state into a vector input to a pre-trained reinforcement learning scheduling network for action optimization, and obtain an initial multi-device linkage action strategy. Step 5: Perform time series smoothing on the initial multi-device linkage action strategy to obtain the corrected multi-device linkage action strategy. Step 6: Map control parameters to the revised multi-device linkage action strategy, generate equipment control commands, and complete the energy-saving scheduling of multi-device linkage in the exhibition hall.

[0007] Secondly, a reinforcement learning-based multi-device linkage energy-saving scheduling system for exhibition halls includes: The acquisition module is used to acquire the first microenvironment parameters located inside the independent microenvironment display case, the second wall heat flux on the surface of the phase change heat storage wall, and the third global environment parameters below the radiant cooling ceiling, to obtain the initial set of environment parameters; The module is used to normalize each parameter in the initial environmental parameter set, map the normalized environmental parameter set to the feature manifold space, calculate the geodesic distance of the parameter vector corresponding to each sampling time in the normalized environmental parameter set in the feature manifold space, and construct the spatial distance matrix. The extraction module is used to perform multi-dimensional scaling transformation on the spatial distance matrix, extract low-dimensional geometric coordinates that preserve the global spatial topology, and obtain the environment state embedding vector. The action optimization module is used to embed the environmental state into the vector input of the pre-trained reinforcement learning scheduling network to optimize the action and obtain the initial multi-device linkage action strategy. The processing module is used to perform time series smoothing on the initial multi-device linkage action strategy to obtain the corrected multi-device linkage action strategy. The control module is used to map control parameters to the modified multi-device linkage action strategy, generate equipment control commands, and complete the energy-saving scheduling of multi-device linkage in the exhibition hall.

[0008] The above-described solution of the present invention has at least the following beneficial effects: This invention effectively quantifies the spatial heat and humidity conduction distance and topological structure between display cases, walls, and ceilings through feature manifold spatial mapping and geodesic distance calculation, overcoming the shortcomings of traditional linear features in representing complex heat and humidity coupling. By utilizing multi-dimensional scaling transformation to extract low-dimensional environmental state embedding vectors, it provides high-fidelity global spatial state representations for reinforcement learning networks, improving the accuracy and rationality of action optimization.

[0009] By combining time series smoothing processing, the fluctuations in action caused by response delay of semiconductor temperature control base and radiant cooling ceiling are effectively mitigated, and stable collaborative linkage of multiple devices is achieved; by generating precise instructions through control parameter mapping, over-adjustment of equipment and cold capacity redundancy are avoided, which significantly reduces the overall energy consumption of the exhibition hall while better ensuring the safe preservation of specimens in the microenvironment display case. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a reinforcement learning-based multi-device linkage energy-saving scheduling method for exhibition halls, provided by an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of a multi-device linkage energy-saving scheduling system for exhibition halls based on reinforcement learning, provided by an embodiment of the present invention.

[0012] Figure 3 This is a schematic diagram of the process for obtaining the environmental state embedding vector provided by an embodiment of the present invention. Detailed Implementation

[0013] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0014] like Figure 1 As shown, embodiments of the present invention propose a multi-device linkage energy-saving scheduling method for exhibition halls based on reinforcement learning. The method includes the following steps: Step 1: Obtain the first microenvironmental parameters located inside the independent microenvironment display case, the second wall heat flux on the surface of the phase change heat storage wall, and the third global environmental parameters below the radiant cooling ceiling to obtain the initial environmental parameter set; Step 2: Normalize each parameter in the initial environmental parameter set, map the normalized environmental parameter set to the feature manifold space, calculate the geodesic distance of the parameter vector corresponding to each sampling time in the normalized environmental parameter set in the feature manifold space, and construct the spatial distance matrix. Step 3: Perform multi-dimensional scaling transformation on the spatial distance matrix to extract low-dimensional geometric coordinates that preserve the global spatial topology, and obtain the environment state embedding vector; Step 4: Embed the environmental state into a vector input to a pre-trained reinforcement learning scheduling network for action optimization, and obtain an initial multi-device linkage action strategy. Step 5: Perform time series smoothing on the initial multi-device linkage action strategy to obtain the corrected multi-device linkage action strategy. Step 6: Map control parameters to the revised multi-device linkage action strategy, generate equipment control commands, and complete the energy-saving scheduling of multi-device linkage in the exhibition hall.

[0015] In a preferred embodiment of the present invention, step 1 specifically includes: A temperature sensor and a humidity sensor are installed at the center of each display case. The accuracy of the temperature sensor is determined by one-tenth of the allowable temperature deviation of the display case, and the accuracy of the humidity sensor is determined by one-tenth of the allowable humidity deviation of the display case. The temperature sensor continuously collects the air temperature value inside the display case at a preset sampling frequency (the sampling frequency range is 1-10Hz, which can be adaptively adjusted according to the intensity of the thermal and humidity parameter fluctuations in the exhibition hall; a high-frequency sampling of 6-10Hz is used under extreme temperature difference and sudden humidity changes in high-altitude and cold conditions, and a low-frequency sampling of 1-5Hz is used under normal and stable exhibition hall conditions, which can accurately capture transient environmental fluctuations while avoiding data redundancy and computing power waste caused by high-frequency sampling). The humidity sensor continuously collects the relative humidity value inside the display case at the same sampling frequency. The temperature value and relative humidity value inside the display case collected at the same sampling time are combined to obtain the first microenvironment parameter. The first microenvironment parameter is a data set containing two components at each sampling time. The first component is the temperature value inside the display case, in degrees Celsius; the second component is the relative humidity value inside the display case, expressed as a percentage.

[0016] Multiple heat flux meters are evenly arranged at preset intervals on the surface of the phase change thermal storage wall facing the inside of the exhibition hall. The preset spacing ranges from 0.8 to 1.2 meters, with 1.0 meter being preferred. Considering the heat and moisture conduction distribution characteristics of the phase change wall in the high-altitude exhibition hall and the requirements of building thermal testing specifications, the lower limit of 0.8 meters can avoid interference from local thermal bridges and local heat transfer anomalies at the corners of the wall, ensuring sufficient sampling density of measuring points and complete capture of subtle differences in the heat flux distribution across the entire wall area. The upper limit of 1.2 meters can avoid redundant measuring point layout, reduce data processing power consumption, and prevent local interference caused by excessive spacing. To address the issue of missed heat transfer characteristics, this method balances the accuracy of heat flux acquisition with operational efficiency. All measuring points are located away from wall cracks, air vents, and areas prone to equipment radiation interference, ensuring the validity of the collected data. A heat flux meter is a sensor that directly measures the heat flux density (i.e., the heat passing through a unit area per unit time) across its mounting surface. Each heat flux meter collects the heat flux value perpendicular to the wall surface at its installation location using the same sampling frequency as the display case sensors. Positive values ​​indicate heat flowing from the wall into the exhibition hall (heat release from the wall), while negative values ​​indicate heat flowing from the exhibition hall into the wall (heat storage from the wall). The arithmetic mean of all heat flux meter readings at the same sampling time is calculated by summing the readings of all heat flux meters at the same moment and dividing by the total number of heat flux meters. This yields the second wall heat flux, a single value at each sampling time, measured in watts per square meter.

[0017] A global temperature and humidity sensor is installed below the radiant cooling ceiling at a preset height from the ceiling surface (the preset sampling height ranges from 1.2 to 1.8 meters, and the sensor is uniformly placed at 1 / 2 of the exhibition hall's net height to effectively avoid interference from the boundary layer airflow near the wall and ensure the accuracy of global environmental parameter acquisition). This global temperature and humidity sensor collects the temperature and dew point temperature values ​​of the air below the ceiling at the same sampling frequency as the aforementioned sensor. The dew point temperature refers to the temperature at which air is cooled to the point where water vapor begins to condense into liquid water under certain humidity conditions. The air temperature and dew point temperature values ​​collected at the same sampling time below the ceiling are combined to obtain the third global environmental parameter.

[0018] Since the sensors inside the display case, the heat flux meters on the wall, and the global sensors under the ceiling may be controlled by different data acquisition terminals, the data collected by each sensor may have slight deviations on the time axis. Therefore, it is necessary to perform timestamp alignment processing on the above three types of parameters. A timestamp (i.e., the precise moment when the data was collected) is automatically added to the data collected by each group of sensors, and then the following alignment rules are executed: the timestamp of the first microenvironment parameter is used as the reference time series. For the second wall heat flux and the third global environment parameter, the moment closest to each moment in the reference time series is found in their own time series (based on the minimum absolute value of the time difference). The collected value corresponding to that moment is used as the aligned value. The maximum allowable time deviation range is preset to ±200ms. If a reference moment cannot match valid data within the deviation range in the time series of another type of parameter (e.g., short-term communication failure of the sensor, data loss), then the reference moment and all corresponding parameters are removed from the set.

[0019] After the above timestamp alignment process, an initial environmental parameter set is obtained. The initial environmental parameter set is a data table, and each row corresponds to a sampling time. The row contains: timestamp, temperature value inside the display case, relative humidity value inside the display case, average heat flux on the wall surface, air temperature value below the ceiling, and dew point temperature value below the ceiling; thus, the initial environmental parameter set is obtained.

[0020] Independent microenvironment display cases: These are independent, enclosed or semi-enclosed cabinets used in exhibition halls to store and display geological specimens. The internal space of these cases exchanges heat and mass (water vapor) with the broader exhibition hall environment through the case walls (usually glass and metal frames). Simply put, a display case is like a small room for the specimen; its internal temperature and humidity are not entirely identical to the temperature and humidity of the broader exhibition hall environment, and there are differences and mutual influences between the two.

[0021] Phase change thermal storage wall: refers to a building wall structure filled with phase change material. The phase change material changes from solid to liquid when the temperature rises to its phase change temperature, absorbing a large amount of heat (thermal storage process); when the temperature drops below the phase change temperature, it changes from liquid to solid, releasing the previously absorbed heat (thermal release process).

[0022] Radiant cooling ceiling: This refers to a type of ceiling that is installed in the ceiling and achieves cooling through radiant heat exchange between the ceiling panel and the surfaces of people and objects in the room. Cold water circulation pipes are embedded inside the ceiling. As the low-temperature water flows through the pipes, it lowers the temperature of the ceiling panel, and the low-temperature ceiling surface absorbs heat from the objects below through thermal radiation.

[0023] A heat flow meter is a sensor that can directly measure the heat flux density (heat passing through a unit area per unit time) through its mounting surface. In this step, the heat flow meter is attached to the surface of the phase change thermal storage wall to sense in real time whether the wall is absorbing or releasing heat, and the intensity of heat transfer.

[0024] Dew point temperature: refers to the temperature at which air, under certain humidity conditions, isobarically cooled to the point where water vapor begins to condense into liquid water (i.e., reaches saturation). When the surface temperature of an object is lower than the dew point temperature of the air, water vapor in the air will condense into water droplets on the object's surface (condensation).

[0025] Timestamp alignment processing: This refers to the operation of matching multi-channel sensor data from different data acquisition terminals to a unified time reference according to their respective recorded time information. The rule used in this step is: using the timestamp of the first microenvironment parameter as the reference, the closest timestamps are found in the time series of the second and third types of parameters for pairing.

[0026] In a preferred embodiment of the present invention, step 2 specifically includes: Because the dimensions and numerical ranges of the parameters in the initial environmental parameter set vary greatly—the internal temperature of the display case may be between ten and twenty degrees Celsius, the relative humidity may be between zero and 100%, and the heat flux of the wall may range from negative tens to positive tens of watts per square meter—directly putting these parameters together for subsequent spatial relationship analysis would cause the parameters with larger value ranges to mask the influence of the parameters with smaller value ranges. Therefore, range normalization is first performed on each parameter separately.

[0027] The specific rules for range normalization are as follows: For each parameter in the initial set of environmental parameters (e.g., the internal temperature of the display case), perform the following operations: Iterate through all the values ​​of the parameter at all sampling times and find the maximum and minimum values; calculate the difference between the maximum and minimum values ​​of the parameter to obtain the range of the parameter; for the original value of the parameter at each sampling time, first subtract the minimum value of the parameter from it, and then divide the resulting difference by the range of the parameter to obtain the normalized value of the parameter at that sampling time; when the range is zero (i.e., the value of the parameter is exactly the same at all times), the normalized value of the parameter at all times is uniformly set to 0.5.

[0028] After range normalization, each sampling time in the initial environmental parameter set corresponds to a parameter vector containing four normalized components: the first component is the normalized display case temperature, the second component is the normalized display case humidity, the third component is the normalized wall heat flux, and the fourth component is the normalized ceiling temperature. This yields the normalized environmental parameter set.

[0029] Based on the normalized environmental parameter set and combined with the spatial location information of each sensor, the spatial coordinates and corresponding thermal and humidity characteristics of locations inside the display case, on the surface of the phase change heat storage wall, and under the radiant cooling ceiling are extracted to construct a three-dimensional spatial thermal and humidity distribution point set. The construction rules of the three-dimensional spatial thermal and humidity distribution point set are as follows: a fixed corner of the exhibition hall is taken as the origin of the spatial rectangular coordinate system, the length direction of the exhibition hall is taken as the first coordinate axis (denoted as the horizontal axis), the width direction is taken as the second coordinate axis (denoted as the vertical axis), and the height direction is taken as the third coordinate axis (denoted as the vertical axis). For each sensor installation location, its three coordinate values ​​in the coordinate system are measured to obtain a three-dimensional spatial coordinate point; the parameter values ​​(temperature, humidity, heat flux, etc.) collected and normalized by the sensor at each sampling time are used as thermal and humidity status labels for the spatial point and attached to the coordinate point. The spatial coordinate points of all sensors and their thermal and humidity status labels are combined to form a three-dimensional spatial thermal and humidity distribution point set.

[0030] Based on a three-dimensional spatial heat and humidity distribution point set, the heat and humidity conduction distance between adjacent spatial locations is calculated to construct a spatial heat and humidity conduction network. The calculation rules for the heat and humidity conduction distance are as follows: For any two spatial locations in the three-dimensional spatial heat and humidity distribution point set, first calculate their Euclidean distance in three-dimensional space, that is, take the squares of the differences in the horizontal, vertical, and vertical coordinates of the two points respectively, sum them, and then take the square root; calculate the difference in heat and humidity state between the two points, that is, take the differences of each component (normalized temperature, normalized humidity, normalized heat flux, normalized ceiling temperature difference) in the heat and humidity state labels of the two points respectively, sum them, and then take the square root; the heat and humidity conduction distance between adjacent spatial locations is determined by both the spatial distance and the difference in heat and humidity state: the heat and humidity conduction distance is equal to the product of the spatial distance and the difference in heat and humidity state; its meaning is: the greater the spatial distance and the greater the difference in heat and humidity state between the two points, the greater the resistance to heat and humidity conduction between the two points, and the less likely they are to directly affect each other. According to the above rules, each point in the three-dimensional spatial heat and humidity distribution point set is taken as a node in the spatial heat and humidity conduction network. An edge is drawn between nodes that have a direct physical connection relationship (e.g., two sensors are located on the same airflow channel, on both sides of the same wall, or on adjacent walls of the same display case). The edge is marked with the heat and humidity conduction distance value calculated according to the above rules. Thus, the spatial heat and humidity conduction network is obtained. It is a weighted undirected graph, where nodes represent spatial points where sensors are installed, edges represent direct physical connections, and the weights on the edges represent heat and humidity conduction distances.

[0031] Based on the spatial heat and moisture conduction network, the cumulative heat and moisture conduction distance between any two spatial locations is calculated, and a spatial distance matrix is ​​constructed. The calculation rule for the cumulative heat and moisture conduction distance is as follows: For any two nodes in the spatial heat and moisture conduction network, along each path connecting them in the network, the heat and moisture conduction distance values ​​of each edge on each path are accumulated one by one, and the path with the smallest accumulated value among all possible paths is taken as the cumulative heat and moisture conduction distance between the two nodes (this rule is similar to selecting the shortest route in map navigation). The above calculation is performed on each pair of nodes in the network, and all calculation results are organized into a matrix. The specific structure of the spatial distance matrix is ​​as follows: the total number of rows in the matrix is ​​equal to the total number of nodes in the network (i.e., the total number of sensor installation locations), and the total number of columns is also equal to the total number of nodes, so it is a square matrix; the i-th row of the matrix corresponds to the i-th node (i.e., the i-th sensor installation location), and the j-th column corresponds to the j-th node. The element at the i-th row and j-th column of the matrix stores the cumulative heat and moisture conduction distance value between the i-th node and the j-th node. Since the resistance from a node to itself is zero (the distance between nodes is zero), all elements on the diagonal of the matrix are zero. Because the resistance from node i to node j is equal to the resistance from node j to node i (the transmission path is reversible), the matrix is ​​symmetric, and the value in the i-th row and j-th column is equal to the value in the j-th row and i-th column; thus, the spatial distance matrix is ​​obtained.

[0032] Normalization processing refers to the process of converting dimensional data into dimensionless data with values ​​that are uniformly within a fixed interval (between zero and one in this embodiment).

[0033] Range normalization is a specific method of normalization. For a set of data, first find the maximum and minimum values, then subtract the minimum value from each data point and divide by the difference between the maximum and minimum values ​​(the range), so that the result falls between zero and one.

[0034] A characteristic manifold space is a new mathematical space to which the original data is projected through a certain mathematical mapping relationship. In this new space, the geometric structure of the data (such as the distance relationship between points) can better reflect the inherent similarity and correlation between the data. In this embodiment, the characteristic manifold space refers to the discrete manifold representation space that uses the shortest path distance along the network edges between nodes in the spatial thermal and moisture conduction network as the approximate geodesic distance of the manifold. The shortest path distance is the sum of the thermal and moisture conduction resistance along each edge. The characteristic manifold space is constructed as follows: the spatial thermal and moisture conduction network is regarded as an undirected weighted graph. Each node in the graph represents a normalized environmental parameter vector corresponding to a sampling time, and the weight of each edge is the thermal and moisture conduction resistance value between the two endpoints. For any two nodes in the undirected weighted graph, all possible paths connecting the two nodes are found, the sum of the edge weights on each path is calculated, and the minimum value is taken as the approximate geodesic distance between the two nodes in the characteristic manifold space. All node pairs are traversed, and the calculated approximate geodesic distances are filled into the matrix to obtain the spatial distance matrix.

[0035] Geodesic distance: refers to the distance between two points measured along the shortest path on the surface of a manifold in the space of a characteristic manifold, rather than a simple straight-line distance; in this embodiment, the cumulative heat and moisture conduction distance is the geodesic distance between two points in the manifold space.

[0036] A spatial heat and moisture conduction network refers to a network structure that abstracts the various sensor measurement locations in an exhibition hall as nodes, the existence of direct heat and moisture exchange channels between these locations as edges, and the ease of conduction along these channels as edge weights. This network can quantify and visualize the flow paths and resistance distribution of heat and moisture in the exhibition hall space.

[0037] In a preferred embodiment of the present invention, step 3 specifically includes: Because the spatial distance matrix has a very high dimension (the number of rows and columns is equal to the total number of sensor nodes), directly using it as input to subsequent reinforcement learning networks would result in an excessively large input dimension, increasing the computational burden and reducing learning efficiency. Therefore, it is necessary to compress the high-dimensional spatial distance matrix to a low dimension through multi-dimensional scaling transformation, while preserving as much as possible the distance relationships between point pairs in the original distance matrix (i.e., the global spatial topology).

[0038] A centered bi-standard transformation is performed on the spatial distance matrix to obtain a centered inner product matrix. The specific rules for the centered bi-standard transformation are as follows: Square each element in the spatial distance matrix, that is, multiply the original value of each position by itself to obtain the squared distance matrix; calculate the row mean vector: for each row of the squared distance matrix, sum the element values ​​of all columns in that row and divide by the total number of columns to obtain the row mean. Arrange all the row means together to form a row mean vector of length equal to the total number of nodes; calculate the column mean vector: for each column of the squared distance matrix, sum the element values ​​of all rows in that column and divide by the total number of rows to obtain the column mean. The column means of all columns are arranged together to form a column mean vector of length equal to the total number of nodes. The overall mean is calculated by summing all element values ​​in the squared distance matrix and dividing by the total number of elements in the matrix. For the element value at row i, column j in the squared distance matrix, its negative value is taken and divided by two. Then, half the row mean of row i is added, half the column mean of column j is added, and finally half the overall mean is subtracted. This result is the element value at row i, column j in the centered inner product matrix.

[0039] The centralized inner product matrix is ​​obtained through the above calculations. The dimension of the centralized inner product matrix is ​​the same as that of the spatial distance matrix, the total number of rows equals the total number of nodes, the total number of columns equals the total number of nodes, and the matrix is ​​symmetric. Eigenvalue decomposition is then performed on the centralized inner product matrix. Eigenvalue decomposition refers to the process of decomposing a square matrix into a set of eigenvalues ​​and corresponding eigenvectors, where each eigenvalue represents the amount of information or energy in the direction of that eigenvector. The calculation rules for eigenvalue decomposition are as follows: Iterative orthogonal decomposition is used, with the following steps: 1. Set an initial vector (dimension equal to the total number of nodes, each component taking a random small value); 2. Multiply this initial vector by the centered inner product matrix to obtain a new vector; 3. Normalize the new vector by calculating the square root of the sum of the squares of its components, then divide each component by this square root to obtain the first eigenvector; 4. Subtract the product of this eigenvector and its transpose from the centered inner product matrix, then multiply by the corresponding Rayleigh quotient (the inner product of the eigenvector and the matrix) to obtain the reduced matrix; 5. Replace the centered inner product matrix with the reduced matrix, repeat the calculation, and sequentially extract subsequent eigenvalues ​​and eigenvectors until the extracted eigenvalues ​​are less than the preset minimum eigenvalue threshold. The preset threshold is set to 1 / 100,000 of the maximum eigenvalue. This threshold is the optimal truncation threshold for dimensionality reduction through multi-dimensional scaling transformation, accurately distinguishing between effective feature modes and noise-redundant modes. Eigenvalues ​​greater than this threshold correspond to the core topological features and effective environmental information of the exhibition hall's thermal and humidity space, while eigenvalues ​​less than this threshold only correspond to minor sensor noise, calculation rounding errors, and invalid weak disturbances, without any actual physical meaning. Using a relative threshold of one ten-thousandth of the maximum eigenvalue instead of a fixed numerical threshold allows for adaptive truncation to accommodate different numbers of sensor nodes in the exhibition hall and different environmental conditions. This thoroughly filters out invalid noise features, avoids dimensional redundancy and overfitting, while retaining more than 95% of the core space's thermal and humidity topological information, balancing algorithm accuracy and computational efficiency. The algorithm stops when the threshold is reached.

[0040] All extracted feature values ​​are arranged in descending order, and the corresponding feature vectors are arranged in the same order. This yields multiple feature values ​​and their corresponding feature vectors arranged in descending order. A predetermined number of target feature values ​​and their corresponding target feature vectors are extracted from the descendingly arranged feature values. The predetermined number is determined as follows: all feature values ​​are summed sequentially from largest to smallest. When the sum reaches a predetermined cumulative ratio threshold (0.95, meaning 95% of the effective feature information is retained), the number of accumulated feature values ​​is the target number. Based on this target number, the top-ranked feature values ​​are extracted as target feature values, and the corresponding feature vectors are extracted as target feature vectors. The target feature vectors are arranged in descending order of their corresponding feature values, from left to right, forming a target feature matrix.

[0041] The square root of each target feature value in the target feature matrix is ​​multiplied sequentially by each component of the corresponding target feature vector. That is, for the k-th target feature value, its square root is calculated first, and then the square root value is multiplied by each component of the k-th target feature vector to obtain a column of coordinate values ​​in the low-dimensional space. The coordinate columns corresponding to all target feature values ​​are combined together to form the low-dimensional geometric coordinates that preserve the global spatial topology. The low-dimensional coordinate vector corresponding to the sampling time is directly used as the environmental state embedding vector at the sampling time.

[0042] Multidimensional scaling transformation is a data analysis method that maps the distance relationships between points in a high-dimensional space to a low-dimensional space as faithfully as possible. Its core objective is to maintain the distance ordering relationship between the original point pairs after dimensionality reduction; point pairs that are far apart in the original space should also be far apart in the low-dimensional space, and point pairs that are close together in the original space should also be close together in the low-dimensional space.

[0043] Centering biscale transformation: This is a key preprocessing step in multidimensional scaling transformation. Its function is to convert the squared distance values ​​in the distance matrix into inner product values ​​in the inner product matrix, and to make the row sum and column sum of the inner product matrix both zero (i.e., centering).

[0044] Eigenvalue decomposition is a mathematical operation that decomposes a square matrix into a set of eigenvalues ​​and corresponding eigenvectors. The magnitude of the eigenvalues ​​reflects the variance of the data along the direction of the eigenvector (i.e., the degree of dispersion of the data in that direction).

[0045] The target feature matrix is ​​a matrix formed by selecting the eigenvectors corresponding to a number of larger eigenvalues ​​from all eigenvectors, arranging them in descending order of eigenvalues. Each column of this matrix is ​​an eigenvector, and the number of columns equals the number of selected eigenvalues ​​(i.e., the target dimension after dimensionality reduction).

[0046] Environmental state embedding vector: This refers to a compact, low-dimensional vector obtained by processing multi-source, high-dimensional, and heterogeneous environmental sensor data within the exhibition hall through a series of processes, including normalization, manifold distance calculation, and multi-dimensional scaling transformation. This vector encodes the thermal and humidity state information of the entire exhibition hall in the most concise form, while preserving the topological relationships between different spatial locations (who is close to whom, who influences whom), and serves as the direct input to reinforcement learning networks.

[0047] In a preferred embodiment of the present invention, step 4 specifically includes: The embedded vector is input into a pre-trained reinforcement learning scheduling network, which autonomously decides what multi-device linkage action strategy should be output in the current environment. The construction process of the reinforcement learning scheduling network is described. This network consists of a policy network and a value network. In this embodiment, the focus is on the structure and training process of the policy network. The policy network adopts an architecture consisting of two linear transformation layers connected in series. The specific structure is as follows: First linear transformation layer: Receives the environment state embedded vector as input. The dimension of this embedded vector is equal to the preset number of target feature values. The first linear transformation layer internally maintains a weight matrix and a bias vector. The number of rows in the weight matrix is ​​equal to the dimension of the embedded vector, and the number of columns is equal to the preset hidden dimension (the hidden dimension is twice the dimension of the embedded vector and rounded down, following the classic network structure design principle of shallow feature extraction; the embedded vector is the compressed low-dimensional core environmental feature, set to twice). The hidden layer with a specific dimension can fully expand the feature representation space, completely uncover the implicit correlations and thermal-wet coupling patterns between environmental features, and avoid the problems of insufficient feature extraction and information loss caused by insufficient dimension. At the same time, it adopts an integer dimension setting with rounding down to adapt to the matrix operation rules of neural networks, eliminates the computational anomalies caused by non-integer dimensions, balances feature representation capability and computational stability, adapts to the feature decoding requirements of low-dimensional embedded vectors in this scheme, and avoids the problems of parameter redundancy, overfitting and computational waste caused by excessively high dimensions. The dimension of the bias vector is equal to the hidden dimension. The processing operation of the first linear transformation layer is as follows: the input embedded vector is multiplied by the weight matrix, that is, the i-th component of the embedded vector is multiplied by the element of the i-th row and j-th column of the weight matrix, the sum of all i is obtained to obtain the j-th intermediate component, and then the j-th intermediate component is added to the j-th component of the bias vector to obtain the first intermediate vector with the dimension of the hidden dimension. Then, a rectified linear function is applied to each component of the first intermediate vector. That is, when the component value is greater than zero, the original value is retained, and when the component value is less than or equal to zero, the component is set to zero. The first intermediate vector after the rectified linear function is used as the output of the first linear transformation layer.

[0048] The second linear transformation layer receives the first intermediate vector after it has been processed by the rectified linear function from the output of the first linear transformation layer. The second linear transformation layer also maintains a weight matrix and a bias vector. The number of rows in the weight matrix is ​​equal to the hidden dimension, and the number of columns is equal to the action dimension (in this embodiment, the action dimension is two, corresponding to the base temperature adjustment compensation value and the ceiling cooling capacity attenuation coefficient, respectively). The dimension of the bias vector is equal to the action dimension. The processing operation of the second linear transformation layer is similar to that of the first layer. The input vector is multiplied by the weight matrix and then added to the bias vector to obtain the action distribution mean vector with a dimension of two.

[0049] In addition to outputting the mean vector of the action distribution, the policy network also maintains a standard deviation vector of the action distribution (also two-dimensional). This standard deviation vector is adaptively updated based on gradient information during training and remains fixed during inference. The policy network outputs the distribution parameters for both action dimensions: for the base temperature control action dimension, its mean is the first component of the mean vector of the action distribution, and its standard deviation is the first component of the standard deviation vector of the action distribution; for the ceiling cooling action dimension, its mean is the second component of the mean vector of the action distribution, and its standard deviation is the second component of the standard deviation vector of the action distribution.

[0050] Based on the aforementioned two-dimensional parameters, the complete construction of two sets of continuous action normal probability distributions is completed step by step. The specific construction process is as follows: First, the core baseline and fluctuation range of the distribution are determined. The mean values ​​corresponding to the two action dimensions are used as the central baseline values ​​of the corresponding probability distributions. These values ​​represent the baseline values ​​for the optimal adjustment actions of the equipment under the current exhibition hall environment conditions. Then, the standard deviation of the corresponding dimension is used to define the size of the action fluctuation range, determining the exploration fluctuation range of the equipment adjustment actions. A continuous action probability interval is constructed. Using the central baseline value as the core, combined with the fluctuation range corresponding to the standard deviation, it is evenly expanded to both sides to form a complete, continuous, and uninterrupted range of executable action values ​​for the equipment, covering all reasonable equipment adjustment actions under this dimension. Corresponding action probability weights are matched. Following the inherent distribution law of the normal distribution, corresponding execution probabilities are assigned to all adjustment actions within the action interval. The closer the adjustment action is to the central baseline, the stronger its ability to adapt to the current working conditions, and the higher the probability of being selected for execution. The farther the adjustment action is from the central baseline, the lower its adaptability to the working conditions, and the lower the probability of being selected for execution. The standard probability distribution curve is formed based on the aforementioned action range and probability matching rules, resulting in a symmetrical bell-shaped probability distribution pattern. The width of the distribution pattern is determined by the standard deviation; the smaller the standard deviation, the more concentrated the distribution curve, tending towards stable optimal actions; the larger the standard deviation, the flatter the distribution curve, indicating stronger action exploration. Through the above steps, the probability distributions of base temperature regulation actions and ceiling cooling capacity attenuation coefficient actions are constructed independently, ultimately integrating to obtain a complete two-dimensional continuous action probability distribution system, providing a compliant and controllable probabilistic basis for subsequent random sampling of equipment actions.

[0051] For the temperature control action distribution of the base, a value is randomly selected from its corresponding normal distribution curve (the selection rule is: the higher the value on the distribution curve, the greater the probability of it being selected), to obtain the base temperature control compensation value; for the cooling action distribution of the ceiling, a value is also randomly selected from its corresponding normal distribution curve to obtain the ceiling cooling attenuation coefficient; these two sampled values ​​are combined into a two-dimensional vector to obtain the continuous action vector. The continuous action vector is then split in dimensions and its features are concatenated: the first component of the continuous action vector is extracted separately as the base temperature control compensation value, which represents the amount of current adjustment that needs to be added or reduced relative to the default operating current of the semiconductor temperature control base under the current environmental conditions (a positive value indicates an increase in heating power, and a negative value indicates an increase in cooling power); the second component of the continuous action vector is extracted separately as the ceiling cooling attenuation coefficient, which is a coefficient between zero and one, indicating that the rated cooling capacity of the radiant cooling ceiling should be multiplied by this coefficient to reduce (attenuate) its actual cooling output under the current environmental conditions. The base temperature adjustment compensation value and the ceiling cooling capacity attenuation coefficient are combined in sequence to obtain the initial multi-device linkage action strategy.

[0052] Before deployment in the actual exhibition hall, the network needs to be pre-trained in a simulation environment. The training process follows these rules: The source and construction of training samples: Using a thermal and humidity environment simulation model of the exhibition hall, a large amount of simulation data is generated under different operating conditions. These conditions include different outdoor temperatures (uniformly sampled from the lowest winter temperature to the highest summer temperature), different visitor flows (uniformly sampled from no visitors during closure to full capacity during peak season), and different initial equipment states. The simulation model is run under each operating condition, recording the environmental parameters at each time step (as states) and the optimal equipment action (as expert action labels) that minimizes total energy consumption and meets sample preservation requirements under that state. Based on the simulation duration and sampling frequency, several state-action pairs are generated as training samples. All training samples are randomly divided into a training set and a validation set according to a preset ratio. The validation set samples do not participate in parameter updates and are only used to evaluate the training effect.

[0053] The core idea of ​​the parameter update strategy is that the difference between the new and old strategies should not be too large during each update. This is equivalent to applying a soft constraint to the strategy's steps, preventing a single update from going too far and causing training instability. Specifically, in each training round, training set samples are input into the current version of the policy network (called the current strategy), which outputs the action probability distribution for each state. Simultaneously, samples of the same state are input into the previous version of the policy network (called the old strategy), which also outputs the action probability distribution. The ratio of the current strategy's output probability to the old strategy's output probability (called the importance weight) is calculated. The dominance function value is then calculated. The dominance function measures how much better an action is compared to the average action in a given state. The calculation of the dominance function uses the generalized dominance estimation method: at each time step, based on the value network's estimate of the state at that moment, the estimate of the state at the next moment, and the immediate reward obtained from the current action, this immediate reward is the core feedback indicator used by this scheduling method to quantitatively evaluate the overall merits of a single-step device scheduling action, providing the sole criterion for iterative optimization of network parameters. The reward consists of two parts: positive incentives and negative constraints. Specifically, it is composed of a negative value of energy consumption cost plus a penalty for operating conditions. The positive incentive is based on the real-time operating energy consumption of the entire temperature control equipment. The lower the energy consumption of the equipment, the stronger the positive incentive effect. The negative constraint is based on the micro-environmental conditions of the exhibition hall and display case. When the temperature and humidity inside the display case deviate from the preset allowable storage range of the specimen, there is a risk of condensation, or the environmental steady state is unbalanced, the penalty constraint is triggered, lowering the overall reward value. The overall evaluation logic is as follows: when the equipment has excellent energy-saving effect, the micro-environment parameters of the display case are stable and compliant, and there are no safety risks, the immediate reward value is higher, indicating that the current scheduling action is better; when the equipment energy consumption is high or the micro-environment conditions are not up to standard, the immediate reward value decreases, indicating that the current scheduling action has defects. The network will optimize the subsequent control strategy accordingly, calculate the time difference error based on the immediate reward, and then gradually reduce the time difference error of several consecutive steps according to the preset discount factor from near to far before weighted accumulation. The preset discount factor ranges from 0.90 to 0.99, and in this embodiment, 0.95 is preferred. This range can balance the weight ratio of immediate reward and long-term return, and adapt to the slow and lagging characteristics of the thermal and humid environment of the exhibition hall. The preferred value can weaken the interference of ineffective gradients in the long term, significantly improve the stability and accuracy of the advantage function estimation, and finally obtain the advantage estimate value at that moment.

[0054] Based on the importance weight and the advantage function value, an alternative loss function is constructed. The design rules of this loss function are as follows: when the advantage value is positive, it encourages increasing the probability of the action, but the importance weight must not exceed the preset upper limit of 1.2 (otherwise it is considered that the policy change is too large and will no longer be encouraged); when the advantage value is negative, it encourages decreasing the probability of the action, but the importance weight must not be lower than the preset lower limit of 0.8; the loss function also includes an entropy reward term, which encourages the policy's action distribution to maintain a certain degree of randomness (the larger the entropy, the flatter the distribution and the stronger the randomness), so as to avoid the policy from getting trapped in local optima too early.

[0055] The gradient of the substitution loss function with respect to each parameter in the weight matrix and bias vector of the policy network is calculated (i.e., the partial derivative of the loss function along the direction of each parameter). The network parameters are iteratively updated based on the gradient information. The overall direction of parameter updates is controlled by the first-order statistical features of the gradient, while an independent learning update step size is matched for each network parameter based on the second-order statistical features of the gradient, so as to achieve adaptive parameter updates. During the parameter update process, the update step size is automatically reduced for dimensions with drastic gradient changes to effectively avoid parameter oscillation and divergence problems; the update step size is automatically increased for dimensions with gentle gradient changes to accelerate the overall convergence speed of the network.

[0056] After each training round, the performance of the current policy network is evaluated using validation set samples, and the average total reward value on the validation set is calculated. The highest historical validation reward value is recorded. If the validation reward value does not improve within a preset number of consecutive evaluation rounds, a preset stagnation round of 50 rounds is set. This round threshold accurately determines the performance saturation state, effectively preventing invalid iterations and avoiding overfitting. Early stopping is triggered to terminate training, and the policy network parameters corresponding to the highest historical validation reward value are saved as the final network weights. A maximum training round limit of 1000 rounds is set to limit the overall training iteration range, avoiding resource waste and excessive runtime issues caused by endless training. If the early stopping condition is not triggered when the training rounds reach this limit, training is forcibly stopped, and the parameters corresponding to the round with the highest validation reward value are selected as the final parameters.

[0057] During the training process described above, the training of the value network and the policy network are carried out simultaneously. The loss function of the value network is the mean squared error between the state value predicted by the value network and the reward value calculated based on the actual cumulative reward. The parameters of the value network are also updated in each round using gradient descent. The weights of the policy network obtained after training constitute the pre-trained reinforcement learning scheduling network, which is deployed in the actual exhibition hall for online inference. During the inference (actual use) phase, the standard deviation of the action distribution of the policy network is maintained at the value saved at the end of training and is no longer updated.

[0058] Reinforcement learning scheduling network: This refers to an artificial neural network that continuously interacts with the environment and optimizes its decision-making strategy based on reward signals from the environment. In this embodiment, the environment refers to the heat and humidity conditions of the exhibition hall, and the reward signal is determined by two factors: energy consumption cost and specimen preservation conditions. The lower the energy consumption and the better the specimen preservation conditions (the more stable the temperature and humidity), the higher the reward value.

[0059] Policy Network: This is a sub-network in the reinforcement learning scheduling network. Its function is to receive the current environmental state information (environment state embedding vector in this embodiment) and output the action to be taken (base temperature adjustment compensation value and ceiling cooling attenuation coefficient in this embodiment). The policy network does not directly give a definite action value, but gives a probability distribution of the action, and then selects a specific action through random sampling. This method retains the possibility of exploration and helps to discover better policies.

[0060] Value network: It is another sub-network in reinforcement learning scheduling network. Its function is to estimate the expected value of the total sum of future cumulative rewards that can be obtained from starting from this state and continuing to execute the current policy, given an environmental state.

[0061] Action probability distribution: refers to the distribution curve of the probability of being selected by the policy network for each possible device adjustment action; in this embodiment, a normal distribution is used, which is uniquely determined by two parameters: mean and standard deviation. The mean represents the most recommended action value, and the standard deviation represents the confidence level of the recommendation. The smaller the standard deviation, the more confident the network is in the action value.

[0062] Base temperature compensation value: refers to the additional current adjustment amount added above the default reference operating current of the semiconductor temperature control base; a positive value indicates further heating on the default basis, and a negative value indicates further cooling on the default basis.

[0063] Ceiling cooling capacity attenuation coefficient: This is a reduction factor multiplied by the rated cooling output (cooling capacity when running at full load) of the radiant cooling ceiling. The value is between zero and one. The closer the coefficient is to one, the closer the actual cooling output of the ceiling is to the full value; the closer the coefficient is to zero, the closer the actual cooling output of the ceiling is to being completely shut off.

[0064] Proximal policy optimization algorithm: This is a widely used policy network training algorithm in reinforcement learning. The core meaning of "proximal" is that each time parameters are updated, the difference between the new policy and the old policy is limited to a certain range to avoid training failure due to excessively large update steps.

[0065] Generalized advantage estimation is a method for estimating how much better each action taken at each time step is than the average. It obtains the advantage value of the current action by integrating reward information from several consecutive future time steps (the reward of the time step further away from the current time contributes less to the current estimate, decaying exponentially).

[0066] In a preferred embodiment of the present invention, step 5 specifically includes: For the base temperature adjustment compensation value, a base temperature adjustment history cache queue is maintained. This queue stores the base temperature adjustment compensation values ​​of the most recent consecutive historical moments with a preset window length. The preset window length is calculated by dividing the response time constant of the semiconductor temperature control base by the control sampling interval and rounding it down. The base response time constant ranges from 3 to 10 seconds, and the corresponding sliding window length ranges from 5 to 20 steps. The greater the device inertia and the longer the response delay, the longer the window length is set, and the greater the smoothing effect. Whenever a new base temperature adjustment compensation value is output, this value is appended to the end of the history cache queue. If the queue length exceeds the preset window length after appending, the earliest value entered at the head of the queue is removed to keep the queue length constant at the window length.

[0067] A moving average is calculated for all values ​​in the historical cache queue. This involves summing all base temperature compensation values ​​in the queue and dividing by the number of values ​​in the queue (i.e., the window length) to obtain a smoothed base temperature compensation value. Since the average is recalculated each time a new value enters and the oldest value is removed, this averaging is called a moving average. Through the moving average process, the sharp fluctuations at a single moment are diluted and smoothed out by the values ​​of multiple moments before and after, resulting in a more gradual change in the final output control quantity. The same processing method is used for the ceiling cooling capacity attenuation coefficient. A historical cache queue for ceiling cooling capacity is maintained (the queue length is also independently determined according to the response time constant of the radiant cooling ceiling). Each newly generated ceiling cooling capacity attenuation coefficient is appended to the end of the queue and the oldest value is removed. Then, a moving average is calculated for all values ​​in the queue to obtain a smoothed ceiling cooling capacity attenuation coefficient. The smoothed base temperature compensation value and the smoothed ceiling cooling capacity attenuation coefficient are recombined in the same order as the initial strategy, i.e., the first component is the smoothed base temperature compensation value and the second component is the smoothed ceiling cooling capacity attenuation coefficient, resulting in the corrected multi-device linkage action strategy.

[0068] Time series smoothing refers to the process of reducing drastic fluctuations in values ​​between adjacent time points by using averaging or filtering methods to make the data change trend smoother.

[0069] Response delay refers to the time elapsed between issuing a command and the controlled device completing the corresponding state change. In this embodiment, the semiconductor temperature control base experiences a response delay of seconds to tens of seconds from receiving a current adjustment command to the actual change in its surface temperature, due to the time required for heat conduction by the semiconductor cooling chip. Similarly, the radiant cooling ceiling experiences an even longer response delay from receiving a water valve opening command to the actual change in the ceiling surface temperature, as the propagation of water temperature changes in the pipeline takes time.

[0070] Moving average: a specific method for smoothing time series data. For each time point, the average value of that time point and several consecutive time points before and after it is taken as the smoothed result for that time point. Because the window "slides" over time, the mean of a time window centered on itself is recalculated at each time point, hence the name moving average.

[0071] Historical cache queue: This refers to a first-in, first-out (FIFO) data storage structure used to store recent action records. New data is placed at the back of the queue, and the oldest data is removed from the front when the queue is full. In this embodiment, the length of the historical cache queue (window length) is determined by the device's response time constant.

[0072] In a preferred embodiment of the present invention, step 6 specifically includes: The temperature compensation value of the smooth base is converted into the base drive current. The specific rules for the physical quantity conversion calculation are as follows: The current-temperature change correspondence of the semiconductor temperature-controlled base is obtained in advance through experimental calibration: Under standard test conditions (fixed ambient temperature and fixed heat dissipation conditions), different magnitudes of DC drive current are applied to the base sequentially, and the temperature change of the base surface after reaching steady state (relative to the temperature when no current is applied) is recorded for each current, forming a set of current-temperature change data pairs; these data pairs are organized into a lookup table, which is a one-dimensional table, where each row corresponds to a temperature change range, and each range stores... The corresponding drive current value for this range is converted according to the following rules in actual use: Add the smooth base temperature adjustment compensation value to the base's default reference current to obtain the target base current value; check whether the target base current value exceeds the maximum allowable drive current of the semiconductor temperature control base (the positive maximum value and the reverse maximum value are determined by the base hardware specifications); if the target base current value is greater than the positive maximum value, truncate it to the positive maximum value; if the target base current value is less than the reverse maximum value (negative value), truncate it to the reverse maximum value; if it is within the allowable range, keep the original value unchanged; use the target base current value after truncation check as the base drive current.

[0073] The cooling capacity attenuation coefficient of the smooth ceiling is converted into the opening degree of the ceiling water valve. The specific rules for the conversion calculation of the physical quantity are as follows: The opening degree of the water valve of the radiant cooling ceiling refers to the opening percentage of the electric regulating valve in the ceiling cold water pipeline. When fully open, the cold water flow is at its maximum (rated cooling capacity), and when fully closed, the cold water flow is zero (no cooling). The opening degree of the ceiling water valve is determined by the cooling capacity attenuation coefficient of the smooth ceiling through the following rules: Determine the rated cooling capacity by obtaining the rated cooling capacity value under standard operating conditions (specific chilled water supply and return temperatures and flow rates) based on the product nameplate parameters of the radiant cooling ceiling; Convert the cooling capacity attenuation coefficient of the smooth ceiling into the target cooling capacity by multiplying the rated cooling capacity by the cooling capacity attenuation coefficient of the smooth ceiling (since the attenuation coefficient is between zero and one, the product is less than or equal to the rated cooling capacity) to obtain the target cooling capacity; Convert the target cooling capacity into the water valve opening degree. Based on the flow-cooling capacity characteristic curve of the ceiling-mounted cold water pipe, the percentage of water valve opening required to generate the target cooling capacity is obtained by referring to a table. The flow-cooling capacity characteristic curve is also obtained through experimental calibration: under standard operating conditions, the water valve opening is adjusted sequentially from zero to 100%, and the actual cooling capacity corresponding to each opening is recorded to form a table of correspondence between opening and cooling capacity. When querying, if the target cooling capacity is exactly equal to the cooling capacity value of a certain record in the table, the corresponding opening is directly taken; if the target cooling capacity falls between two records, linear interpolation is used, that is, the two endpoint values ​​are weighted and averaged according to the proportional relationship between the target value and the two endpoints to obtain the water valve opening. The percentage of water valve opening is converted into a control signal that the water valve actuator can recognize (usually a current signal of four to twenty milliamps or a voltage signal of zero to ten volts) to obtain the ceiling-mounted water valve opening.

[0074] The base drive current and ceiling water valve opening are standardized and sequentially packaged into commands to complete the closed-loop energy-saving scheduling of multiple devices in the exhibition hall. The specific execution process is as follows: Data is packaged according to the device's compatible communication protocol. For the semiconductor temperature control base, the calculated accurate base drive current value is standardized and packaged according to the device's serial bus communication protocol. The device address code, function command code, current value data segment, and data check code are filled in sequentially to generate a complete first device control command frame, ensuring that the command format fully matches the identification specifications of the temperature control base controller. For the electric regulating valve of the radiant cooling ceiling, the calibrated ceiling water valve opening percentage value is segmented and packaged according to the valve's dedicated communication transmission protocol. The regulating valve device address, opening adjustment function code, opening value field, and check field are filled in sequentially to generate a compliant second device control command frame. Unified timing marking and data verification: For two sets of packaged control command frames, the current precise timestamp is synchronously bound to mark the time of command generation and issuance, avoiding issues such as command timing disorder and data asynchrony among multiple devices; the integrity of the command is verified through a verification field, eliminating invalid commands with missing data or abnormal format, ensuring the accuracy of transmitted data. Hierarchical timing issuance and device execution: Following the scheduling logic of prioritizing micro-environment control and assisting global environment adaptation, the control command frame for the cabinet semiconductor temperature control base is issued first, and the control command frame for the ceiling electric regulating valve is issued only after the base control command transmission is completed, realizing orderly linkage and control of multiple devices. The equipment parsing and execution process involves each device controller receiving the corresponding instruction frame, automatically parsing the valid control values, discarding redundant verification fields and timestamp fields, and driving the semiconductor temperature control base to adjust the heating or cooling power, driving the electric regulating valve to adjust the opening of the cold water pipeline, and changing the cold water flow and cooling output of the ceiling. Through the above standardized and time-sequential instruction encapsulation and issuance process, precise and coordinated control of the micro-environment temperature control of the display case and the global ceiling cooling of the exhibition hall is achieved, ultimately completing the entire process of multi-device linkage energy-saving scheduling in the high-altitude exhibition hall.

[0075] Control parameter mapping refers to the process of converting abstract control quantities (such as compensation values, attenuation coefficients, etc.) used internally by the algorithm into physical control quantities (such as current, valve opening, etc.) that the controlled equipment can directly execute.

[0076] Base drive current: This refers to the direct current that drives the thermoelectric cooler in the semiconductor temperature control base. When the current flows in one direction, one side of the thermoelectric cooler heats up while the other side cools down (Peltier effect); when the current flows in the opposite direction, the heating and cooling sides are interchanged. The magnitude of the current determines the cooling and heating power.

[0077] Ceiling water valve opening: refers to the percentage of opening of the electric regulating valve in the chilled water pipeline of the radiant cooling ceiling. Fully open (100%) means that chilled water flows through the ceiling pipeline at the maximum flow rate, the ceiling surface temperature is the lowest, and the cooling capacity is the maximum; fully closed (zero) means that chilled water does not flow, and the ceiling is basically not cooling.

[0078] Command encapsulation refers to the process of packaging calculated control values ​​into standard control command frames (containing fields such as target address, command type, data content, and checksum) according to the communication protocol format followed by the target device, so that they can be sent to the device controller through the communication network. In simpler terms, command encapsulation is about standardizing the calculated control values ​​into a standardized command format that can be accurately recognized and transmitted by the device controller, according to the device's established communication specifications. This ensures reliable transmission and accurate execution of control commands, avoiding problems such as data corruption and parsing failures.

[0079] A lookup table is a data mapping table pre-established through experimental calibration or simulation calculations. The table stores the mapping relationship between different input conditions and corresponding output results. During runtime, the output results are quickly retrieved by looking up the table, avoiding repetitive and complex calculations. In this embodiment, the lookup table is used for quick lookup of the current-temperature change correspondence and the valve opening-cooling capacity correspondence.

[0080] like Figure 2 As shown, embodiments of the present invention also provide a reinforcement learning-based multi-device linkage energy-saving scheduling system for exhibition halls, including: The acquisition module is used to acquire the first microenvironment parameters located inside the independent microenvironment display case, the second wall heat flux on the surface of the phase change heat storage wall, and the third global environment parameters below the radiant cooling ceiling, to obtain the initial set of environment parameters; The module is used to normalize each parameter in the initial environmental parameter set, map the normalized environmental parameter set to the feature manifold space, calculate the geodesic distance of the parameter vector corresponding to each sampling time in the normalized environmental parameter set in the feature manifold space, and construct the spatial distance matrix. The extraction module is used to perform multi-dimensional scaling transformation on the spatial distance matrix, extract low-dimensional geometric coordinates that preserve the global spatial topology, and obtain the environment state embedding vector. The action optimization module is used to embed the environmental state into the vector input of the pre-trained reinforcement learning scheduling network to optimize the action and obtain the initial multi-device linkage action strategy. The processing module is used to perform time series smoothing on the initial multi-device linkage action strategy to obtain the corrected multi-device linkage action strategy. The control module is used to map control parameters to the modified multi-device linkage action strategy, generate equipment control commands, and complete the energy-saving scheduling of multi-device linkage in the exhibition hall.

[0081] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0082] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A reinforcement learning-based multi-device linkage energy-saving scheduling method for exhibition halls, characterized in that, The method includes: Step 1: Obtain the first microenvironmental parameters located inside the independent microenvironment display case, the second wall heat flux on the surface of the phase change heat storage wall, and the third global environmental parameters below the radiant cooling ceiling to obtain the initial environmental parameter set; Step 2: Normalize each parameter in the initial environmental parameter set, map the normalized environmental parameter set to the feature manifold space, calculate the geodesic distance of the parameter vector corresponding to each sampling time in the normalized environmental parameter set in the feature manifold space, and construct the spatial distance matrix. Step 3: Perform multi-dimensional scaling transformation on the spatial distance matrix to extract low-dimensional geometric coordinates that preserve the global spatial topology, and obtain the environment state embedding vector; Step 4: Embed the environmental state into a vector input to a pre-trained reinforcement learning scheduling network to optimize actions and obtain an initial multi-device linkage action strategy. Step 5: Perform time series smoothing on the initial multi-device linkage action strategy to obtain the corrected multi-device linkage action strategy. Step 6: Map control parameters to the revised multi-device linkage action strategy, generate equipment control commands, and complete the energy-saving scheduling of multi-device linkage in the exhibition hall.

2. The energy-saving scheduling method for multi-device linkage in exhibition halls based on reinforcement learning according to claim 1, characterized in that, Step 1: Obtain the first microenvironmental parameters located inside the independent microenvironment display case, the second wall heat flux on the surface of the phase change heat storage wall, and the third global environmental parameters below the radiant cooling ceiling to obtain the initial environmental parameter set, including: The first microenvironment parameter is obtained by collecting the internal temperature and humidity of the display case from the temperature sensor located inside the independent microenvironment display case; The second wall heat flux is obtained by collecting the heat flux on the wall surface from the heat flux meter located on the surface of the phase change heat storage wall. The third global environmental parameter is obtained by collecting the air temperature and dew point temperature below the ceiling from the global temperature and humidity sensor located below the radiant cooling ceiling. The initial set of environmental parameters is obtained by aligning the first micro-environment parameter, the second wall heat flux, and the third global environment parameter with timestamps.

3. The energy-saving scheduling method for multi-device linkage in exhibition halls based on reinforcement learning according to claim 2, characterized in that, Step 2: Normalize each parameter in the initial environmental parameter set, map the normalized environmental parameter set to the feature manifold space, calculate the geodesic distance of the parameter vector corresponding to each sampling time in the normalized environmental parameter set in the feature manifold space, and construct a spatial distance matrix, including: Based on the first micro-environmental parameter, the second wall heat flux and the third global environmental parameter in the initial environmental parameter set, range normalization is performed respectively to obtain a normalized environmental parameter set. Based on the normalized environmental parameter set, the spatial coordinates and corresponding thermal and humidity characteristics of the locations inside the independent microenvironment display case, on the surface of the phase change heat storage wall, and under the radiant cooling ceiling are extracted to construct a three-dimensional spatial thermal and humidity distribution point set. Based on the three-dimensional spatial heat and humidity distribution point set, calculate the heat and humidity conduction distance between adjacent spatial locations and construct a spatial heat and humidity conduction network; Based on the spatial heat and moisture conduction network, the cumulative heat and moisture conduction distance between any two spatial locations is calculated to obtain the spatial distance matrix.

4. The energy-saving scheduling method for multi-device linkage in exhibition halls based on reinforcement learning according to claim 3, characterized in that, Step 3: Perform a multi-dimensional scaling transformation on the spatial distance matrix to extract low-dimensional geometric coordinates that preserve the global spatial topology, obtaining the environment state embedding vector, including: calculating the square value of each element in the spatial distance matrix according to the spatial distance matrix and performing centralization double standard transformation to obtain a centralized inner product matrix; performing eigenvalue decomposition calculation according to the centralized inner product matrix to obtain a plurality of eigenvalues arranged in descending order and corresponding eigenvectors; extracting a preset number of target eigenvalues and corresponding target eigenvectors from the plurality of eigenvalues arranged in descending order to obtain a target eigenmatrix; performing product calculation on the square root of the target eigenvalue in the target eigenmatrix and the corresponding target eigenvector to obtain low-dimensional geometric coordinates that maintain the global spatial topology, and obtaining an environment state embedding vector through combination.

5. The energy-saving scheduling method for multi-device linkage in exhibition halls based on reinforcement learning according to claim 4, characterized in that, Step 4: inputting the environment state embedding vector into a pre-trained reinforcement learning scheduling network for action optimization to obtain an initial multi-device linkage action strategy, including: inputting the environment state embedding vector into the policy network of the pre-trained reinforcement learning scheduling network for feature mapping to obtain an action probability distribution; performing random sampling according to the action probability distribution to obtain a continuous action vector; performing dimension splitting according to the continuous action vector to separate a base temperature adjustment compensation value and a ceiling cooling capacity attenuation coefficient; performing feature splicing according to the base temperature adjustment compensation value and the ceiling cooling capacity attenuation coefficient to obtain the initial multi-device linkage action strategy.

6. The energy-saving scheduling method for multi-device linkage in exhibition halls based on reinforcement learning according to claim 5, characterized in that, inputting the environment state embedding vector into the policy network of the pre-trained reinforcement learning scheduling network for feature mapping to obtain an action probability distribution, including: inputting the environment state embedding vector into the hidden layer of the policy network for feature extraction processing to obtain a state hidden feature vector; inputting the state hidden feature vector into the output layer of the policy network for action dimension mapping to obtain the distribution parameter of the base temperature adjustment action corresponding to the semiconductor temperature control base and the distribution parameter of the ceiling cooling capacity action corresponding to the radiant cooling ceiling; performing probability distribution construction processing according to the distribution parameter of the base temperature adjustment action and the distribution parameter of the ceiling cooling capacity action to obtain the action probability distribution.

7. The energy-saving scheduling method for multi-device linkage in exhibition halls based on reinforcement learning according to claim 6, characterized in that, Step 5: performing time series smoothing processing on the initial multi-device linkage action strategy to obtain a corrected multi-device linkage action strategy, including: constructing a base temperature adjustment time series according to the base temperature adjustment compensation value and performing moving average calculation to obtain a smoothed base temperature adjustment compensation value; constructing a ceiling cooling capacity time series according to the ceiling cooling capacity attenuation coefficient and performing moving average calculation to obtain a smoothed ceiling cooling capacity attenuation coefficient; reorganizing according to the smoothed base temperature adjustment compensation value and the smoothed ceiling cooling capacity attenuation coefficient to obtain the corrected multi-device linkage action strategy.

8. The energy-saving scheduling method for multi-equipment linkage in exhibition halls based on reinforcement learning according to claim 7, characterized in that, Step 6: performing control parameter mapping on the corrected multi-device linkage action strategy to generate a device control instruction, and completing the energy-saving scheduling of multi-device linkage in the exhibition hall, including: performing physical quantity conversion calculation according to the smoothed base temperature adjustment compensation value to obtain a base driving current; performing physical quantity conversion calculation according to the smoothed ceiling cooling capacity attenuation coefficient to obtain a ceiling water valve opening; performing instruction encapsulation according to the base driving current and the ceiling water valve opening to obtain the device control instruction, and completing the energy-saving scheduling of multi-device linkage in the exhibition hall.

9. A multi-device linkage energy-saving scheduling system for exhibition halls based on reinforcement learning, wherein the system implements the method as described in any one of claims 1 to 8, characterized in that, including: The acquisition module is used to acquire the first microenvironment parameters located inside the independent microenvironment display case, the second wall heat flux on the surface of the phase change heat storage wall, and the third global environment parameters below the radiant cooling ceiling, to obtain the initial set of environment parameters; The module is used to normalize each parameter in the initial environmental parameter set, map the normalized environmental parameter set to the feature manifold space, calculate the geodesic distance of the parameter vector corresponding to each sampling time in the normalized environmental parameter set in the feature manifold space, and construct the spatial distance matrix. The extraction module is used to perform multi-dimensional scaling transformation on the spatial distance matrix, extract low-dimensional geometric coordinates that preserve the global spatial topology, and obtain the environment state embedding vector. The action optimization module is used to embed the environmental state into the vector input of the pre-trained reinforcement learning scheduling network to optimize the action and obtain the initial multi-device linkage action strategy. The processing module is used to perform time series smoothing on the initial multi-device linkage action strategy to obtain the corrected multi-device linkage action strategy. The control module is used to map control parameters to the modified multi-device linkage action strategy, generate equipment control commands, and complete the energy-saving scheduling of multi-device linkage in the exhibition hall.

Images