Intelligent control method for building air conditioner based on deep reinforcement learning

Abstract

The application relates to the technical field of intelligent air conditioner control, in particular to a building air conditioner intelligent control method based on deep reinforcement learning. The method comprises the following steps: combining environmental temperature, surface heat exchange coefficient, heat exchange area and specific heat capacity to execute heat accumulation and generate virtual heat storage; combining building topology to construct a space network and extract a global heat state; combining a carbon emission penalty value extracted from a power grid carbon emission factor, the virtual heat storage and the global heat state to splice and generate a composite state vector, inputting a decision network with an environmental static bias item to output an initial instruction; extracting a reconstruction error of the composite state vector to generate a state confidence, performing amplitude clipping to generate a physical attenuation coefficient, performing amplitude scaling on the initial instruction to generate a safety control instruction, and delivering the safety control instruction to a bottom mechanism. The application forms a mechanism coupling a physical parameter, an external time sequence and a safety constraint, promotes the safety control instruction to keep coherent in action logic, and gives the equipment a more stable operation rhythm.

Description

Technical Field

[0001] This invention relates to the field of intelligent air conditioning control technology, specifically to an intelligent control method for building air conditioning based on deep reinforcement learning. Background Technology

[0002] Intelligent scheduling and energy consumption control of air conditioning systems in large building spaces is an important technological application area. Existing technologies generally adopt control methods based on preset temperature control rules combined with conventional data-driven algorithms. That is, by monitoring the changes in the temperature scalar of the indoor and outdoor environment, when it deviates from the target threshold, the underlying air conditioning units are automatically calculated and issued adjustment commands in a pre-established single control loop. This is currently the mainstream technical means to achieve temperature control in building environments.

[0003] However, existing technical solutions have significant shortcomings in global thermal state assessment and ensuring the safe operation of underlying structures. Existing state perception methods are usually limited to isolated single areas, mainly relying on real-time feedback of air temperature. This local perspective of control logic ignores the heat transfer inertia of physical building complexes during long-term operation, as well as the energy interaction interference caused by the structural connectivity between adjacent physical spaces. The scheduling actions of existing systems are often reactive, only activating high-power regulation after the temperature exceeds the limit, causing equipment to frequently operate at full load during peak energy consumption periods, wasting the natural energy buffer potential of the building structure. In addition, in the output stage of control commands, the action parameters generated by existing algorithms usually lack reasonable engineering boundaries, often directly mapping extreme control actions, which can easily cause violent oscillations in underlying fans or water valves. In summary, due to the lack of in-depth analysis of the building's heat storage potential and the necessary industrial fault-tolerant operation mechanisms, existing control systems are prone to frequent overload of underlying equipment and even damage caused by violent oscillations.

[0004] To address this, a building air conditioning intelligent control method based on deep reinforcement learning is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a building air conditioning intelligent control method based on deep reinforcement learning to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a building air conditioning intelligent control method based on deep reinforcement learning, comprising: Obtain ambient temperature, surface heat transfer coefficient, heat transfer area, and specific heat capacity; input the ambient temperature and surface heat transfer coefficient into the flux model to extract convective heat flux, and then combine it with the heat transfer area and specific heat capacity to perform heat accumulation to extract sensible heat capacity; input the sensible heat capacity into the mapping space to generate virtual heat storage. Obtain the building topology and structural thermal conductivity; set the virtual heat storage as node features and the structural thermal conductivity as edge weights, and construct a spatial network by combining the building topology; perform neighborhood aggregation of node features along edge weights within the spatial network to extract the global thermal state; Obtain the power grid carbon emission factor; input the power grid carbon emission factor into the time series model to extract the carbon emission penalty value, and concatenate it with the virtual heat storage and global thermal state to generate a composite state vector; extract the environmental static bias term; input the composite state vector and the environmental static bias term into the decision network to output the initial command; The composite state vector is input into the variational autoencoder to perform reconstruction and extract the reconstruction error; the reconstruction error is input into the confidence evaluation channel to perform uncertainty quantization to generate the state confidence; the state confidence is introduced into the boundary channel to perform amplitude limiting and clipping to generate the physical attenuation coefficient; the initial command is combined with the physical attenuation coefficient to perform amplitude scaling to generate the safety control command.

[0007] Preferably, the specific process for generating the sensible heat capacity includes: obtaining a reference temperature, a capacity cutoff boundary, wall thickness, material density, and time step; in the flux model, performing a difference operation between the ambient temperature and the reference temperature to generate a temperature deviation value; performing a product operation between the surface heat transfer coefficient and the temperature deviation value to obtain the surface heat flux; extracting convective heat flux by multiplying the surface heat flux and the heat transfer area; extracting the structural volume base by multiplying the heat transfer area and the wall thickness; generating a solid mass base by multiplying the structural volume base and the material density; obtaining a heat capacity reference base by multiplying the solid mass base and the specific heat capacity; extracting the absolute amount of heat exchange by multiplying the convective heat flux and the time step; introducing the absolute amount of heat exchange into the heat capacity reference base to perform a quotient operation to generate a temperature change; converting the temperature change into a normalized thermal state; and introducing the normalized thermal state into the capacity cutoff boundary to perform boundary constraints to generate the sensible heat capacity.

[0008] Preferably, the specific generation process of the virtual heat storage includes: inputting the sensible heat capacity into the mapping space to obtain the capacity mapping weight parameter; performing a dot product operation on the sensible heat capacity and the capacity mapping weight parameter to generate a weighted capacity; performing a nonlinear transformation on the weighted capacity to extract the capacity activation value; obtaining the environmental penalty compensation coefficient; performing an aggregation operation on the capacity activation value and the environmental penalty compensation coefficient to obtain the compensated capacity; performing a scaling transformation on the compensated capacity to generate a scaled capacity; obtaining the upper limit threshold and the lower limit threshold of capacity saturation; combining the scaled capacity with the upper limit threshold and the lower limit threshold of capacity saturation to perform extreme value restriction to obtain the legal interval capacity; and performing a smoothing and denoising operation on the legal interval capacity to generate virtual heat storage.

[0009] Preferably, the specific generation process of the global thermal state includes: parsing the building topology to extract vertex coordinates and connection positions; mapping the virtual heat storage to the vertex coordinates to generate an initial node feature map; allocating the structural thermal conductivity to the connection positions to generate an initial edge feature map; combining the initial node feature map and the initial edge feature map to construct an original spatial network; performing neighbor node feature transfer on the initial node feature map and the initial edge feature map in the original spatial network to generate neighborhood aggregation features; performing coefficient allocation operation on the neighborhood aggregation features to obtain attention enhancement features; using the attention enhancement features and the initial node feature map to perform concatenation to generate cascaded fused node features; and performing global pooling operation on the cascaded fused node features to extract the global thermal state.

[0010] Preferably, the specific generation process of the composite state vector includes: obtaining a benchmark carbon emission reference index; performing sequence analysis on the power grid carbon emission factor in conjunction with the benchmark carbon emission reference index based on a time-series model to generate time-series carbon emission features; performing numerical amplification processing on the time-series carbon emission features to obtain a carbon emission penalty value; performing dimensional mapping on the virtual heat storage and global thermal state respectively to extract dimensionality-reduced heat storage features and dimensionality-reduced thermal state features; performing dimensional concatenation operation on the carbon emission penalty value, dimensionality-reduced heat storage features, and dimensionality-reduced thermal state features to generate a joint state feature; and performing cross-correlation operation on the joint state feature to generate a composite state vector.

[0011] Preferably, the specific generation process of the initial instruction includes: acquiring historical operating parameters and benchmark design parameters; performing deviation quantization calculations on the historical operating parameters and benchmark design parameters to generate static difference features; performing numerical transformation on the static difference features to obtain environmental static bias terms; inputting the composite state vector combined with the environmental static bias terms into the state input layer within the decision network to extract action hidden states; using the value evaluation layer to perform expectation estimation on the action hidden states to obtain value evaluation coefficients; combining the value evaluation coefficients to perform distribution sampling on the action hidden states to generate action space distribution; parsing the action space distribution to output the initial instruction.

[0012] Preferably, the specific process for generating the state confidence score includes: introducing the composite state vector into the encoding channel of the variational autoencoder to perform distribution prediction, generating mean features and variance features; combining the mean features and variance features to perform sampling operations and extract latent variables in the latent space; performing inverse mapping on the latent variables in the latent space to obtain the reconstructed state vector; performing distance calculation on the reconstructed state vector and the composite state vector to extract the reconstruction error; performing distribution evaluation estimation on the reconstruction error to generate an uncertainty index; and performing inverse numerical transformation on the uncertainty index to generate the state confidence score.

[0013] Preferably, the specific generation process of the safety control command includes: obtaining a preset upper confidence threshold, a lower confidence threshold, and a safety baseline command; performing interval verification and filtering on the state confidence combined with the upper and lower confidence thresholds to extract a valid confidence level; performing a scaling transformation on the valid confidence level to generate a physical attenuation coefficient; performing a difference operation on the initial command and the safety baseline command to generate a command bias amplitude value; combining the command bias amplitude value and the physical attenuation coefficient, generating a safety bias amplitude value through multiplication and reduction operations; performing a command recombination and superposition operation on the safety baseline command and the safety bias amplitude value to extract a safety control command; and converting the safety control command into a physical control signal and sending it to the underlying mechanism.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By setting virtual heat storage as node features and structural thermal conductivity as edge weights, and combining this with building topology to construct a spatial network, neighborhood aggregation is performed along the edge weights to extract the global thermal state, thus analyzing the building's inherent heat storage potential and spatial heat transfer logic. This allows the building's inherent thermal inertia to act as an energy buffer, enabling air conditioning equipment to smoothly manage heat during environmental changes. This alleviates the blindness of equipment response to sudden local temperature changes and avoids frequent overloads of underlying mechanisms induced by transient command jumps, reducing the operational pressure on electromechanical components from the source.

[0015] 2. The composite state vector is input into the variational autoencoder to extract the reconstruction error. After quantizing and generating the state confidence, a boundary channel is introduced to limit and prune the data to generate a physical attenuation coefficient. The initial command is then scaled to generate a safety control command. This transforms the uncertainty of the environmental state into a flexible constraint for the issued actions. When encountering extreme working conditions or external disturbances, the control model can sense the decision risk and implement a smooth degradation of the command, keeping the equipment action within the safety envelope. This avoids the violent oscillations caused by model decision deviation and provides physical protection for the underlying mechanism.

[0016] 3. The carbon emission factor of the power grid is input into the time series model to extract the carbon emission penalty value, which is then concatenated with the virtual heat storage and global thermal state to generate a composite state vector. This vector is then combined with the environmental static bias term to input the decision network and output the initial command. By cascading the dynamic time series characteristics with the spatial physical state, the decision framework establishes a global perspective that takes into account multiple dimensions. This suppresses the full-load output behavior that is prone to occur when equipment pursues a single short-term goal, smooths out the overall operating curve, reduces the hidden dangers of frequent start-stop and large-scale operating condition switching, and maintains the self-consistent stability of energy dispatch.

[0017] 4. Based on the building topology, the global thermal state is extracted and concatenated with carbon emission penalty values ​​to form a composite state vector for outputting initial commands. This vector is then input into a variational autoencoder to evaluate state confidence, forming a mechanism that couples physical parameters, external timing, and safety constraints, avoiding biased decisions based on a single dimension. The physical state, carbon emission trends, and underlying limiting protection form a mutually constraining closed loop, ensuring consistent action logic for safety control commands. This resolves abrupt switching of electromechanical components caused by interplay between different strategy dimensions, giving the equipment a more robust operating rhythm. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent building air conditioning control method based on deep reinforcement learning proposed in an embodiment of this invention application; Figure 2 This is a flowchart of the spatial thermal topology feature extraction based on physical mechanisms proposed in an embodiment of this invention application; Figure 3 This is a flowchart of the composite state decision-making and VAE dual-channel amplitude limiting control proposed in an embodiment of this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1-3 The intelligent control method for building air conditioning based on deep reinforcement learning provided by this invention has the following specific steps: Obtain ambient temperature, surface heat transfer coefficient, heat transfer area, and specific heat capacity; input the ambient temperature and surface heat transfer coefficient into the flux model to extract convective heat flux, and then combine it with the heat transfer area and specific heat capacity to perform heat accumulation to extract sensible heat capacity; input the sensible heat capacity into the mapping space to generate virtual heat storage. Obtain the building topology and structural thermal conductivity; set the virtual heat storage as node features and the structural thermal conductivity as edge weights, and construct a spatial network by combining the building topology; perform neighborhood aggregation of node features along edge weights within the spatial network to extract the global thermal state; Obtain the power grid carbon emission factor; input the power grid carbon emission factor into the time series model to extract the carbon emission penalty value, and concatenate it with the virtual heat storage and global thermal state to generate a composite state vector; extract the environmental static bias term; input the composite state vector and the environmental static bias term into the decision network to output the initial command; The composite state vector is input into the variational autoencoder to perform reconstruction and extract the reconstruction error; the reconstruction error is input into the confidence evaluation channel to perform uncertainty quantization to generate the state confidence; the state confidence is introduced into the boundary channel to perform amplitude limiting and clipping to generate the physical attenuation coefficient; the initial command is combined with the physical attenuation coefficient to perform amplitude scaling to generate the safety control command.

[0021] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0022] Example 1 This application discloses a building air conditioning intelligent control method based on deep reinforcement learning, see embodiments. Figure 1 The specific steps proposed in this invention include: S1. Obtaining ambient temperature, surface heat transfer coefficient, heat transfer area, and specific heat capacity; inputting the ambient temperature and surface heat transfer coefficient into a flux model to extract convective heat flux, and then combining it with the heat transfer area and specific heat capacity to perform heat accumulation to extract sensible heat capacity; inputting the sensible heat capacity into a mapping space to generate virtual heat storage; S2. Obtaining building topology and structural thermal conductivity; setting the virtual heat storage as node features and the structural thermal conductivity as edge weights, and constructing a spatial network in combination with the building topology; performing neighborhood aggregation along the edge weights of the node features within the spatial network to extract the global thermal state; S3. Obtaining grid carbon... Carbon emission factor; input the power grid carbon emission factor into the time series model to extract the carbon emission penalty value, and concatenate it with the virtual heat storage and global thermal state to generate a composite state vector; extract the environmental static bias term; input the composite state vector and the environmental static bias term into the decision network to output the initial command; S4, input the composite state vector into the variational autoencoder to perform reconstruction and extract the reconstruction error; input the reconstruction error into the confidence assessment channel to perform uncertainty quantization to generate the state confidence; introduce the state confidence into the boundary channel to perform amplitude limiting and clipping to generate the physical attenuation coefficient; combine the initial command with the physical attenuation coefficient to perform amplitude scaling to generate the safety control command.

[0023] Further, the ambient temperature, surface heat transfer coefficient, heat transfer area, and specific heat capacity are obtained; the ambient temperature, combined with the surface heat transfer coefficient, is input into the flux model to extract convective heat flux, and then combined with the heat transfer area and specific heat capacity to perform heat accumulation to extract sensible heat capacity; the sensible heat capacity is input into the mapping space to generate virtual heat storage; corresponding to step S1 above; the specific implementation process includes: The process involves obtaining a reference temperature, capacity cutoff boundary, wall thickness, material density, and time step. In the flux model, the difference between the ambient temperature and the reference temperature is calculated to generate a temperature deviation. The surface heat flux is obtained by multiplying the surface heat transfer coefficient with the temperature deviation. Convective heat flux is extracted by multiplying the surface heat flux with the heat transfer area. The structural volume base is extracted by multiplying the heat transfer area with the wall thickness. The structural volume base is multiplied with the material density to generate a solid mass base. The solid mass base is multiplied with the specific heat capacity to obtain a heat capacity reference base. The absolute amount of heat exchange is extracted by multiplying the convective heat flux with the time step. The absolute amount of heat exchange is introduced into the heat capacity reference base for quotient calculation to generate a temperature change. The temperature change is converted into a normalized thermal state. The normalized thermal state is introduced into the capacity cutoff boundary for boundary constraints to generate sensible heat capacity.

[0024] Specifically, the process of generating sensible heat capacity is as follows: The main control unit reads the memory-mapped file to obtain the reference temperature, capacity cutoff boundary, wall thickness, material density, and time step. The controller reads the value from the target demand temperature register address set on the indoor temperature control panel and assigns it as the reference temperature. The main control unit retrieves the extreme value of the maximum daily temperature range from the past 5 years of meteorological records in the local climate history database and directly sets it as the capacity cutoff boundary value. The processor parses the 3D bounding box data in the architectural drawing file, calculates the geometric coordinate difference between the inner and outer walls, and extracts the wall thickness. The main control unit directly reads the communication polling cycle configuration word at the bottom layer of the edge controller and maps it completely to the time step.

[0025] For material density, specific heat capacity, and structural thermal conductivity, the processor extracts material tags from the building information model and performs hash matching in the standard thermophysical property dictionary to extract the corresponding attribute values.

[0026] In this embodiment, after extracting the standard structural thermal conductivity value through dictionary matching, the Euclidean spatial distance is calculated by calling the coordinates of adjacent vertices in the building topology. A distance attenuation division operation is then performed to generate the structural thermal conductivity considering the spatial span. Specifically, the processor calls the parsed vertex coordinates of the building topology to calculate the Euclidean straight-line distance between adjacent nodes in the three-dimensional coordinate system. To establish a unified spatial reference, the processor traverses and extracts the node spacing with the smallest value in the current network, setting it as the standard grid reference distance. Based on this, a division operation is performed with the standard grid reference distance as the numerator and the actual Euclidean spatial distance between the two nodes as the denominator to calculate a universally applicable distance attenuation multiplier. Subsequently, the processor multiplies the initial structural thermal conductivity value obtained through dictionary hash matching with this distance attenuation multiplier to obtain the corrected value after spatial span reduction. This process, combined with existing topological coordinates, endows the thermal conductivity with a distance attenuation attribute, making the allocation of spatial network edge weights more consistent with real physical laws.

[0027] Within the flux model, the arithmetic logic unit performs a difference operation between the ambient temperature and the reference temperature to generate a temperature deviation value. The processor acquires the flow velocity per second transmitted by the wind speed sensor on the exterior wall surface, and extracts the corresponding surface heat transfer coefficient using linear interpolation in a lookup function. The wind speed and heat transfer coefficient mapping reference data in the lookup function are derived from the empirical table of external surface convective heat transfer coefficients in GB 50176.

[0028] In this embodiment, after extracting the basic surface heat transfer coefficient through linear interpolation, an exponentially weighted moving average filtering operation is performed, combining the system time step with historical values ​​from the previous cycle, to output a smoothed surface heat transfer coefficient. Specifically, the processor extracts the historical surface heat transfer coefficient from the previous running cycle in the data buffer, obtains a preset time constant parameter, calculates the negative of the ratio of the current system time step parameter to the preset time constant parameter, uses this negative number as the exponent, and performs an exponential operation on the natural constant e. The result of this operation is set as a smoothing decay factor with a value between 0 and 1, used to reflect the data decay characteristics in the time dimension. Subsequently, the processor multiplies the current surface heat transfer coefficient extracted by linear interpolation with this smoothing decay factor to obtain the retention weight of the current environmental state; simultaneously, it multiplies the difference between the historical surface heat transfer coefficient and (1 - smoothing decay factor) to calculate the historical weight reflecting the system's thermal inertia. Finally, the above-mentioned retention weight and historical weight are added together to obtain the smoothed final value. The time constant parameter is selected as follows: read the hardware datasheet of the underlying edge controller, obtain its maximum I / O communication delay clock period, and strictly set the time constant parameter to 10 times that delay clock period. 3 The above process filters out instantaneous changes in lookup values ​​caused by gusts, ensuring the stability of the surface heat transfer coefficient input into the flux model.

[0029] The surface heat flux is obtained by multiplying the surface heat transfer coefficient and the temperature deviation. Convective heat flux is extracted by multiplying the surface heat flux and the heat transfer area. The multiplier extracts the structural volume base by multiplying the heat transfer area and the wall thickness. The structural volume base is then multiplied by the material density to generate the solid mass base. Finally, the solid mass base is multiplied by the specific heat capacity to obtain the heat capacity reference base.

[0030] During physical transfer, the temperature change is calculated using the following formula:

[0031] In this formula, Indicates the amount of temperature change. Indicates the surface heat transfer coefficient. Indicates ambient temperature. Indicates the reference temperature, and A represents the heat transfer area. This indicates the time step, and 'e' represents the wall thickness. Indicates the density of the material. This represents the specific heat capacity. The numerator of the formula is calculated by multiplying the convective heat flux by the time step, representing the absolute amount of heat exchange; the denominator represents the baseline heat capacity. These two are then used to quotient the temperature change for the current period. It should be noted that since both the extraction of the convective heat flux and the extraction of the structural volume baseline use the same surface heat transfer area for calculation, the two heat transfer areas cancel each other out in the quotient calculation of the temperature change. This reflects the engineering assumption of this embodiment: that the unit area thermal response characteristics of a homogeneous enclosure structure are consistent. The introduction of the heat transfer area is to ensure the comparability of the two intermediate quantities—surface heat flux and structural volume baseline—in terms of physical dimensions (heat exchange quantity and solid mass) in the early stages of the calculation.

[0032] The temperature change is then converted into a normalized thermal state. The processor extracts the highest and lowest absolute temperatures from the local weather station's records over the past 365 days, calculates the difference between them, and sets it as the empirical maximum temperature difference constant. The arithmetic logic unit performs a quotient operation on the temperature change and the empirical maximum temperature difference constant; the resulting normalized thermal state is then introduced into the capacity truncation boundary to perform boundary constraints, and the constrained scalar is output as the sensible heat capacity.

[0033] By introducing time step and thermal capacity baseline to calculate temperature change, feature extraction strictly follows the thermodynamic closed loop, avoiding parameter distortion caused by numerical splicing and providing solid underlying support.

[0034] The sensible heat capacity is input into the mapping space to obtain the capacity mapping weight parameters; a dot product operation is performed on the sensible heat capacity and the capacity mapping weight parameters to generate the weighted capacity; a nonlinear transformation is performed on the weighted capacity to extract the capacity activation value; an environmental penalty compensation coefficient is obtained; an aggregation operation is performed on the capacity activation value and the environmental penalty compensation coefficient to obtain the compensated capacity; a scaling transformation is performed on the compensated capacity to generate the scaled capacity; the upper and lower capacity saturation thresholds are obtained; the scaled capacity is combined with the upper and lower capacity saturation thresholds to perform extreme value constraints to obtain the legal interval capacity; a smoothing and denoising operation is performed on the legal interval capacity to generate virtual heat storage.

[0035] Specifically, the process of generating virtual heat storage is as follows: The main control unit instantiates a fully connected feedforward neural network layer with 64 neurons in memory as the mapping space. The model initialization program calls a pseudo-random number generator to generate a standard normal distribution floating-point number sequence with a mean of 0 and a standard deviation of 0.01, which is then filled into the weight matrix register to form the initial capacity mapping weight parameters. These initial capacity mapping weight parameters serve as learnable parameters for the front-end feature extraction layer during subsequent model training, participating in end-to-end backpropagation and gradient updates.

[0036] The matrix multiplication instruction set performs a dimension broadcast copy of the sensible heat capacity, which is a one-dimensional scalar, aligning its size with the capacity mapping weight parameter vector of size 1×64. Then, it performs an element-wise multiplication operation on the two to produce a weighted capacity containing 64 feature dimensions. A nonlinear transformation operation is performed on the weighted capacity, and a modified linear unit activation function is used to extract the maximum value of each element in the weighted capacity vector after comparing it with 0, generating the capacity activation value.

[0037] The edge gateway monitors relative humidity sensors installed on the exterior of the building. When the measured value of the relative humidity sensor is greater than the historical monthly average, the variable address of the environmental penalty compensation coefficient is assigned the reciprocal of the historical monthly average; otherwise, it is assigned 0. An element-wise aggregation addition operation is performed using the capacity activation value and the environmental penalty compensation coefficient to obtain the compensated capacity.

[0038] The compensated capacity is scaled to generate a scaled capacity, and then the capacity of the legal interval is obtained through a formula, which is described below:

[0039] In this formula, Indicates the capacity of the valid interval. Indicates the capacity after compensation. Indicates the scaling factor. This represents the lower threshold of capacity saturation. This indicates the upper limit threshold for capacity saturation. This represents the extremum constraint function. The network layer hyperparameters are set as follows: the scale factor is set to 0.05, the lower threshold for capacity saturation is set to -1, and the upper threshold for capacity saturation is set to 1.

[0040] The processor performs a Fast Fourier Transform (FFT) on the raw ambient temperature time-series signal continuously collected by the ambient temperature sensor. It then uses a low-pass filter with a preset cutoff frequency to extract low-frequency components and calculate the signal power. It also extracts high-frequency components above the cutoff frequency, calculates their variance as white noise power, calculates the ratio of white noise power to signal power, and uses the negative of this ratio as the exponent. It then performs an exponential operation on the natural constant e, and sets the resulting value, strictly within the (0, 1] interval, as the smoothing factor. The preset cutoff frequency is selected as follows: The measured temperature step response curve of the target building envelope is obtained, and the thermodynamic time constant required to reach two-thirds of the steady-state temperature difference is extracted. Its reciprocal is used as the fundamental frequency. Strictly according to the Nyquist sampling theorem, half of this fundamental frequency is taken as the physical upper limit threshold of the preset cutoff frequency.

[0041] The virtual heat storage at the current time step is equal to the product of the legal interval capacity at the current time step and the smoothing factor, plus the product of the virtual heat storage at the previous time step and the difference between the constant 1 and the smoothing factor, so that the virtual heat storage remains smooth and continuous.

[0042] By using interval pruning and smoothing channels to perform extreme value limiting and noise reduction on the capacity parameter, numerical spikes caused by external disturbances are filtered out, preventing sudden noise from being transmitted to the network.

[0043] Further, obtain the building topology and structural thermal conductivity; set the virtual heat storage as node features and the structural thermal conductivity as edge weights, and construct a spatial network based on the building topology; perform neighborhood aggregation along the edge weights within the spatial network to extract the global thermal state; this corresponds to step S2 above; see [link to relevant documentation]. Figure 2 The specific implementation process includes: The process involves analyzing the building topology to extract vertex coordinates and connection positions; mapping virtual heat storage to vertex coordinates to generate initial node feature maps; allocating structural thermal conductivity to connection positions to generate initial edge feature maps; combining the initial node feature maps with the initial edge feature maps to construct the original spatial network; performing neighbor node feature transfer on the initial node feature maps and initial edge feature maps in the original spatial network to generate neighborhood aggregation features; performing coefficient allocation operations on the neighborhood aggregation features to obtain attention-enhanced features; concatenating the attention-enhanced features with the initial node feature maps to generate cascaded fused node features; and performing global pooling on the cascaded fused node features to extract the global thermal state.

[0044] Specifically, the generation process of the global thermal state is as follows: The processor reads the vertex list structure in the video memory, parses the building topology to extract the vertex coordinates representing the geometric center of each independent temperature-controlled hot zone, and extracts the adjacency relationship between adjacent hot zones sharing a physical partition wall as the line position representing the connectivity relationship.

[0045] For each independent temperature control zone in the building topology, the ambient temperature, external wall heat exchange area and specific heat capacity of the temperature control zone are obtained independently according to the aforementioned flux model and mapping space processing flow. Based on this, the virtual heat storage of the temperature control zone is calculated separately.

[0046] The generated virtual heat storage is then mapped to the corresponding vertex coordinates to generate an initial node feature map matrix; the structural thermal conductivity data is placed into a sparse matrix and assigned to the connection positions to generate an initial edge feature map; the initial node feature map and the initial edge feature map are combined to construct an original spatial network containing an adjacency list in memory.

[0047] In the original spatial network, node features are combined with edge features to perform neighborhood aggregation within the graph attention network layer, generating neighborhood aggregation features. The calculation of attention weights is described using the graph attention mechanism formula as follows:

[0048] In this formula, Indicates attention weights, The weight vector representing the attention mechanism. The linear mapping matrix representing the shared weights. Indicates the characteristics of the central node. Representing the characteristics of adjacent nodes, This represents the characteristics of nodes in the set of neighboring nodes. This represents the thermal conductivity of the line structure connecting the central node and adjacent nodes. This represents the thermal conductivity of the connection structure between the central node and the nodes in the set of neighboring nodes. This represents the set of neighboring nodes of the central node. This indicates a concatenation operation, exp represents the exponential function, and LeakyReLU represents the modified linear unit activation function.

[0049] The network layer hyperparameters are set as follows: The main control unit performs cross-validation based on the memory bandwidth limitation of the edge computing device and the channel requirements of the node feature map, and fixes the dimension of the linear mapping matrix of shared weights to 32 dimensions. The weight vector w of the attention mechanism is set as a one-dimensional learnable parameter matrix with a dimension length of 65.

[0050] The above formula is used to perform coefficient allocation operation on the neighborhood aggregation features, extract the attention score of each neighbor node, and perform scalar product with the corresponding node features to obtain the attention enhancement features.

[0051] To avoid excessive smoothing of features across layers, attention-enhanced features and initial node feature maps are concatenated along the feature dimension to generate concatenated fused node features. The main control unit performs global pooling on the concatenated fused node features, calculates the arithmetic mean along the node dimension, compresses the graph-level data into a one-dimensional tensor, and extracts the global hot state.

[0052] By leveraging graph network feature propagation to perform neighborhood aggregation and attention allocation, the mutual influence of heat transfer between adjacent regions is quantified, enabling the extracted global thermal state to encompass the overall thermal inertia and preventing control commands from making blind and excessive responses to single-point local temperature anomalies.

[0053] Further, obtain the power grid carbon emission factor; input the power grid carbon emission factor into the time series model to extract the carbon emission penalty value, and concatenate it with the virtual heat storage and global thermal state to generate a composite state vector; extract the environmental static bias term; input the composite state vector and the environmental static bias term into the decision network to output the initial command; corresponding to step S3 above; see [link to relevant documentation]. Figure 3 The specific implementation process includes: Obtain benchmark carbon emission reference indicators; perform sequence analysis on the grid carbon emission factor and benchmark carbon emission reference indicators based on the time series model to generate time series carbon emission characteristics; perform numerical amplification processing on the time series carbon emission characteristics to obtain carbon emission penalty values; perform dimensional mapping on virtual heat storage and global thermal state respectively to extract dimensionality-reduced heat storage characteristics and dimensionality-reduced thermal state characteristics; perform dimensional cascading operation on the carbon emission penalty value, dimensionality-reduced heat storage characteristics and dimensionality-reduced thermal state characteristics to generate joint state characteristics; perform cross-correlation operation on the joint state characteristics to generate composite state vectors.

[0054] Specifically, the process of generating the composite state vector is as follows: The database connector queries publicly available data tables provided by the power grid dispatch center, extracts the power grid carbon emission factors for the past 24 hours corresponding to the time window of the Long Short-Term Memory (LSTM) network described below, and calculates the arithmetic mean as a benchmark carbon emission reference index. Based on the LSTM network, the control unit performs sequence analysis on the current power grid carbon emission factors in conjunction with the benchmark carbon emission reference index to generate time-series carbon emission features. The network layer hyperparameters are set as follows: the time window length of the LSTM network is set to 24 hours; furthermore, the LSTM network contains one hidden layer with a fixed number of 64 hidden unit neurons; and the final output time-series carbon emission features are 64-dimensional one-dimensional tensors.

[0055] The edge gateway obtains the ratio of the maximum to minimum carbon emission intensity of fossil fuel power generation in the past 7 days, as publicly available from the local power grid dispatch center, and sets this ratio as a sensitivity multiplier constant. Each value in the time-series carbon emission feature tensor is multiplied by this sensitivity multiplier constant to obtain a carbon emission penalty value.

[0056] Dimensional mapping is performed on both the virtual heat storage and the global thermal state. The computing node uses two independent single-layer feedforward networks to map these two values ​​into tensors, extracting dimensionality-reduced heat storage features and dimensionality-reduced thermal state features. The network layer hyperparameters are set as follows: the length of the dimensionality-reduced tensor is fixed at 128 dimensions. The processor projects the carbon emission penalty value onto the 128-dimensional dimension along the memory channel dimension, and then treats it, along with the dimensionality-reduced heat storage and thermal state features, as independent sequence nodes. A cascaded concatenation operation is performed along the sequence length dimension to generate a joint state feature matrix containing three sequence nodes with a feature dimension of 128 dimensions.

[0057] A cross-association operation is performed on the joint state features to generate a composite state vector. A self-attention mechanism is used to assign association weights, and the calculation is described by the following formula:

[0058] In this formula, Represents a composite state vector. Represents the query matrix. Represents the key matrix. Represents a value matrix, represents the scaling factor of the key matrix, and Softmax represents the normalized exponential function.

[0059] The joint state features are projected into the query matrix, key matrix, and value matrix after passing through the fully connected layer. The network layer hyperparameters are set as follows: the scaling factor for the key matrix is ​​set to the square root of the number of channels in the key matrix. The tensor output after the cross-association operation retains the dimensional structure of the three sequence nodes. The processor flattens this output tensor along the sequence node dimension, reshaping it into a one-dimensional vector of length 384 (i.e., the product of the three sequence nodes and the 128-dimensional feature dimension) as the final composite state vector. This composite state vector is passed as a common input to the downstream reinforcement learning decision network and VAE reconstruction network. During the model backpropagation training phase, the fully connected layer that generates the query matrix, key matrix, and value matrix serves as a shared feature fusion hub. Its gradient comes from the joint backpropagation of the downstream two branches: it simultaneously receives the gradient derived from the policy update loss output from the value evaluation layer of the decision network, and the gradient derived from the reconstruction error of the variational autoencoder. These two gradients are automatically merged and accumulated at the output of the cross-association module, jointly driving the weight updates of the query matrix, key matrix, and value matrix.

[0060] By cascading carbon emission penalties with reduced physical states, dynamic environmental costs are injected into the decision-making and observation space, prompting the decision-making process to weigh long-term operational rhythms and mitigating the large-scale full-load output behavior induced by simply pursuing short-term temperature control targets.

[0061] The process involves: acquiring historical operating parameters and baseline design parameters; performing deviation quantification on the historical operating parameters and baseline design parameters to generate static difference features; performing numerical transformation on the static difference features to obtain environmental static bias terms; inputting the composite state vector and environmental static bias terms into the state input layer within the decision network to extract action hidden states; using the value assessment layer to perform expectation estimation on the action hidden states to obtain value assessment coefficients; combining the value assessment coefficients to perform distribution sampling on the action hidden states to generate action space distribution; and analyzing the action space distribution to output initial instructions.

[0062] Specifically, the initial instruction generation process is as follows: The processor queries the local database to obtain historical indoor operating temperature parameters and baseline design temperature parameters from the design drawings. It performs deviation quantification calculations on the historical operating temperature parameters and the baseline design temperature parameters, i.e., performs a subtraction operation in the arithmetic unit to generate static difference characteristics. The static difference characteristics are then numerically transformed and normalized by dividing by a temperature difference constant to obtain the environmental static bias term. The microprocessor analyzes the acceptable absolute temperature drift limit when the PMV index is 0 in the annex to the international ISO 7730 standard, extracts the index value, and sets it as the temperature difference constant.

[0063] The composite state vector, combined with a static environmental bias term, is input into the state input layer of the policy network. The network hyperparameters are set as follows: the policy network contains three hidden layers with fixed numbers of neurons: 256, 128, and 64. The network forward computes to extract the action hidden state; subsequently, the value evaluation layer extracts this action hidden state and performs expectation estimation on its internal nodes to calculate the value evaluation coefficient.

[0064] Combining the value assessment coefficients, the policy network performs near-end policy optimization parameter updates. The control unit extracts the product of the ratio of the probability of actions from the old and new policies and the advantage function value as the first comparison value; the ratio of the probability of actions from the old and new policies is truncated using the truncation range hyperparameter, and then the truncated output value is multiplied by the advantage function value as the second comparison value; the minimum value between the first and second comparison values ​​is extracted as the policy update loss value. The network layer hyperparameters are set as follows: the truncation range hyperparameter is based on the Lipshitz continuity constant of the system state transition matrix and the maximum gradient norm of the current network action output layer, substituted into the theoretical lower bound formula of KL divergence, and the theoretical threshold that ensures the single-step policy divergence does not exceed the bound is solved in reverse.

[0065] The network outputs the mean and variance parameters of a multivariate Gaussian distribution, performs distribution sampling to generate an action space distribution, maps the distribution to a defined numerical range using a hyperbolic tangent function, and outputs initial instructions in one-dimensional floating-point format with a numerical range between -1 and 1.

[0066] By converting the deviation between historical operation and design benchmark into a static environmental bias term, customized anchor points are provided for action distribution sampling, the influence of building characteristics on the decision network is calibrated, and the initial analysis of instruction distribution closely matches the actual operating boundary of electromechanical components.

[0067] Further, the composite state vector is input into the variational autoencoder to perform reconstruction and extract the reconstruction error; the reconstruction error is input into the confidence evaluation channel to perform uncertainty quantization and generate a state confidence level; the state confidence level is introduced into the boundary channel to perform amplitude limiting and clipping to generate a physical attenuation coefficient; the initial command is combined with the physical attenuation coefficient to perform amplitude scaling and generate a safety control command; corresponding to the above S4 step; the specific implementation process includes: The composite state vector is introduced into the encoding channel of the variational autoencoder to perform distribution prediction, generating mean and variance features. Sampling operations are performed by combining the mean and variance features to extract latent variables in the latent space. Inverse mapping is performed on the latent variables in the latent space to obtain the reconstructed state vector. Distance calculation is performed between the reconstructed state vector and the composite state vector to extract the reconstruction error. The distribution evaluation and estimation of the reconstruction error are performed to generate an uncertainty index. Inverse numerical transformation is performed on the uncertainty index to generate the state confidence score.

[0068] Specifically, the process of generating state confidence is as follows: The composite state vector is simultaneously fed into the encoding channel of the variational autoencoder to perform distribution prediction. The encoding channel contains two cascaded fully connected layers. First, the input 384-dimensional composite state vector is mapped to an intermediate hidden layer containing 128 neurons. Then, the dimensionality is reduced by two parallel fully connected output layers, resulting in mean and variance features with a fixed dimension of 64. Sampling operations are performed on the mean and variance features to extract latent variables in the latent space. The nodes are then mapped to the log-variance features using the natural exponential function, transforming them into strictly non-negative variance features. Subsequently, the square root operation is performed on the variance features to be reconstructed to extract the standard deviation tensor. The element-wise product of this standard deviation tensor and the random noise matrix is ​​then calculated, and the result is added to the mean feature. The sum of the outputs is used as the latent variable in the latent space. The network layer hyperparameters are set as follows: the dimension of the latent variables in the latent space of the variational autoencoder is set to 64 dimensions.

[0069] The latent variables in the latent space are input into the decoding channel for inverse mapping, and the network reconstructs the dimensions of the input data to obtain the reconstructed state vector. Distance calculation is performed between the reconstructed state vector and the original composite state vector to extract the reconstruction error. The tensor operation unit calculates the difference between a single element of the original composite state vector and the corresponding element of the reconstructed state vector, and squares this difference. The squared values ​​under all dimensions are summed, and finally, the sum is divided by the number of dimensions of the vector to extract the reconstruction error.

[0070] To evaluate and estimate the distribution of reconstruction error, the processor subtracts the mean error of historical data from the reconstruction error of the current period and divides it by the standard deviation of historical error to generate an uncertainty index.

[0071] In this embodiment, the uncertainty index is subjected to absolute value calculation to unify the deviation direction, and a natural logarithm function is used for nonlinear transformation to generate a final uncertainty index that smooths out extreme abnormal peaks. Specifically, the processor takes the absolute value of the initially generated uncertainty index, eliminates the difference in mathematical signs between positive and negative deviations, and converts it into an absolute deviation scalar reflecting the pure fluctuation amplitude. To meet the domain requirements of subsequent logarithmic operations and prevent computational anomalies, the processor adds a constant 1 to this absolute deviation scalar to construct the input true value for nonlinear mapping. Subsequently, the built-in natural logarithm function is called to perform a transformation operation on this input true value. Utilizing the convergence characteristics of the logarithmic curve, the linear amplification deviation that may occur under extremely rare conditions is mapped to a gradually increasing logarithmic value. Finally, the logarithmically compressed numerical output is used as the uncertainty index participating in the subsequent reverse numerical transformation channel. The above process uses logarithmic mapping to nonlinearly suppress extreme reconstruction deviations, avoiding irrational sharp drops in subsequent confidence levels caused by transient anomalies.

[0072] An inverse numerical transformation is performed on the uncertainty index to generate a state confidence score. The processor statistically analyzes all reconstruction error records from the historical normal operating condition verification set during the initial debugging phase of the device, extracts the 99th quantile floating-point error value, and sequentially performs the same standardization operation (subtracting the historical error mean and dividing by the standard deviation), absolute value increment operation, and natural logarithmic nonlinear transformation on this value, extracting the boundary eigenvalue at a uniform scale. The reciprocal of this boundary eigenvalue is then calculated and set as the penalty sensitivity coefficient. The arithmetic unit calculates the product of the penalty sensitivity coefficient and the uncertainty index, adds a constant 1 to the product as the divisor, performs division with the constant 1 as the dividend, and outputs the quotient as the state confidence score.

[0073] By relying on decoding and reconstruction operations to extract deviation distance and uncertainty indicators, the degree of deviation of environmental feature distribution is objectively quantified in an unsupervised state and transformed into state confidence, which is conducive to identifying unknown extreme working conditions and perceiving potential risks.

[0074] The system acquires preset upper and lower confidence thresholds and safety baseline commands; it combines the state confidence with the upper and lower confidence thresholds to perform interval verification and filtering to extract the confidence; it performs a scaling transformation on the confidence to generate a physical attenuation coefficient; it performs a subtraction operation between the initial command and the safety baseline command to generate a command bias amplitude value; it combines the command bias amplitude value and the physical attenuation coefficient to generate a safety bias amplitude value through multiplication and reduction operations; it performs command recombination and superposition operations on the safety baseline command and the safety bias amplitude value to extract the safety control command; and it converts the safety control command into a physical control signal and sends it to the underlying mechanism.

[0075] Specifically, the generation process of security control instructions is as follows: Obtain the preset upper confidence threshold, lower confidence threshold, and safety baseline command. The safety baseline command is read directly from the PID control register in the device's underlying controller. The processor retrieves the confidence statistical distribution of the model output under historical normal operating condition verification set, extracts the 95th quantile value and sets it as the upper confidence threshold; it also extracts the 5th quantile value and sets it as the lower confidence threshold.

[0076] The state confidence is combined with upper and lower confidence thresholds for interval verification and filtering: if the confidence is higher than the upper confidence threshold, the upper threshold is used; if the confidence is between the upper and lower thresholds, the original value is retained; if the state confidence is lower than the lower confidence threshold, the processor extracts the historical valid confidence from the previous control cycle and obtains a preset decay step constant. The selection method is as follows: based on the mechanical torque overload protection threshold fixed in the servo driver hardware manual, and combined with the time step of the current control cycle, the maximum allowable opening increment per step without triggering hardware overload protection is calculated, and this increment is directly used as the decay step constant; the product of the current state confidence and the decay step constant, plus the product of the historical valid confidence and the difference between (1 minus the decay step constant), is used as the valid confidence extracted after performing exponential smooth transition. The valid confidence is then divided by the upper confidence threshold to perform a scaling transformation, generating a physical decay coefficient. The processor extracts the upper and lower limit parameters of the device's actuator opening, and linearly maps the initial instruction to a percentage opening format consistent with the safety baseline instruction's dimension through an affine transformation. Then, it performs a difference operation between the unified initial instruction and the safety baseline instruction to generate the instruction bias magnitude value. Combining the instruction bias magnitude value and the physical attenuation coefficient, a multiplication reduction operation is used to generate the safety bias magnitude value.

[0077] The safety control instruction is extracted by combining the safety baseline instruction with the safety bias amplitude value and performing instruction recombination and superposition operations. The logic control unit extracts the difference between the initial instruction and the safety baseline instruction, calculates the product of this difference and the physical attenuation coefficient, and finally performs an addition operation on the product with the safety baseline instruction, outputting the sum as the safety control instruction.

[0078] The calculated safety control commands are converted into physical control signals and sent to the underlying mechanisms. This stage relies on the building automation and control network protocol stack. The communication layer network parameters are set as follows: the communication port is mapped to the default setting 47808; the write command is specified to the 8th level of the priority array. The microprocessor formats the floating-point format safety control commands into application layer messages at the IP network layer and writes them into the current value attribute of the corresponding actuator's simulated output object.

[0079] After the edge-side data controller extracts and updates the current value of the buffer, it converts the percentage floating-point number into an analog electrical signal through a high-precision 12-bit digital-to-analog converter chip with opto-isolation protection. The electrical signal is transmitted to the physical device terminal through shielded twisted-pair cable to drive the fan inverter, or to the stepper motor driver where it is converted into pulses by V / F conversion to drive the chilled water valve stepper motor, thus completing the execution of the physical control closed loop.

[0080] By subtracting the initial command from the safety baseline and using the attenuation coefficient to shrink the action bias, the command can be forced to smoothly revert to the vicinity of the physical baseline, thus avoiding abrupt switching or damage to the underlying mechanism due to receiving out-of-limit commands.

[0081] The architectural details, data processing flow, and specific training mechanism of the deep learning network model in this embodiment are described below: When constructing the training sample set for this deep learning network model, the underlying data mainly comes from the historical operation logs of the target building and synchronously recorded meteorological data; the input features specifically cover ambient temperature, physical temperature measurement data, air conditioning operating conditions, and carbon emission factors. As for the label column, it can be obtained from historical manually adjusted action sequences and overall energy consumption statistics; after preprocessing, the relevant training sample set will be divided into training set, validation set, and test set, which will respectively participate in gradient calculation, early stopping monitoring to prevent model overfitting, and final evaluation of the model's generalization ability.

[0082] The entire model training process is divided into two parallel pathways: unsupervised reconstruction and policy optimization. The side-channel monitoring module, responsible for feature reconstruction, needs to simultaneously ensure reconstruction accuracy and constrain the distribution of latent variables; therefore, its joint loss function is a combination of mean squared error reconstruction loss and relative entropy. The core module for decision control employs a round-based experience replay mechanism. To encourage the model to explore actions in the early stages of training, the joint loss function of this core module not only includes the policy loss with truncation and the mean squared error loss of state value, but also adds an additional policy cross-entropy reward.

[0083] The reward signal for the strategy optimization branch is composed of indoor temperature deviation and equipment energy consumption: the single-step reward is the negative value of the absolute value of the deviation between the historical manual debugging action sequence and the current output initial command, plus the negative value of the overall energy consumption statistics within the corresponding control cycle, and the normalized negative energy consumption value obtained by dividing the overall energy consumption statistics within the corresponding control cycle by the rated energy consumption benchmark value. The two are then weighted and summed according to preset weights to obtain the immediate reward for that step. In this embodiment, the rated full-load power on the chiller unit nameplate is extracted as the energy consumption normalization denominator, and the maximum mechanical stroke limit of the valve actuator is extracted as the action deviation normalization denominator. The preset weight ratio of energy consumption and action deviation is determined by the reciprocal ratio of their normalization coefficients. The preferred value for the action deviation weight is 0.6, and the preferred value for the energy consumption weight is 0.4. The historical data recorded in the training label column is used to calculate the above rewards step by step, rather than as an independent supervision target. The value assessment layer fits the cumulative return of the reward sequence based on the discount, and the difference between this and the actual cumulative return constitutes the advantage function value.

[0084] After entering the backpropagation optimization stage, gradient descent is performed using the Adam optimizer. In this stage, the joint loss of the policy optimization branch and the reconstruction loss of the unsupervised reconstruction branch are added to form the global total loss. The master control unit calculates the gradient of this global total loss with respect to all learnable parameters and backpropagates the gradient to the frontmost mapping space using the chain rule. The Adam optimizer is used to synchronously update the capacity mapping weight parameters end-to-end, enabling the originally randomly initialized parameter space to adaptively adjust under the target's drive. This dynamically maps the sensible heat capacity, which has a definite physical meaning, into a nonlinear latent variable that includes time lag and heat transfer inertia characteristics.

[0085] Regarding parameter configuration, the global initial learning rate is set to 10. -4It also employs a cosine annealing decay strategy; the batch size is set to 256. The discount factor specific to reinforcement learning is set to 0.99, the generalized advantage estimation parameter is set to 0.95, and the maximum number of offline training rounds is set to 1000. However, if the validation set loss does not decrease for 50 consecutive rounds, an early stopping mechanism will be triggered immediately, at which point the final model weight parameters will be directly fixed and output for subsequent real-world inference and deployment.

[0086] This invention provides an intelligent control method for building air conditioning based on deep reinforcement learning. It combines virtual heat storage and building topology to construct a spatial network, thereby extracting the global thermal state. This network is then combined with the time-series characteristics of the power grid to form a composite state, which is then fed into the decision network. Furthermore, a variational autoencoder is introduced, which quantifies a confidence level through reconstruction error, thereby generating a corresponding attenuation coefficient to scale the initial command. This method directly couples spatial heat transfer mechanisms, dynamic carbon emission costs, and safety constraints of actions at the underlying level, mitigating the problem of blind response of equipment to local abrupt changes. Even under extremely harsh operating conditions, it allows control commands to be smoothly degraded, ultimately protecting electromechanical components from damage caused by frequent overloads and large oscillations.

[0087] Example 2 This embodiment describes the execution flow of a building air conditioning intelligent control method based on deep reinforcement learning in a real physical environment.

[0088] Edge computing devices receive outdoor ambient temperature via an industrial serial communication bus. To accurately assess the thermal inertia of the building's exterior walls, the computing unit combines the surface heat transfer coefficient with specific physical properties. Here, the thickness of the load-bearing wall is taken as 0.25 meters, the material density is 2200 kg / m³, and the specific heat capacity corresponds to 900 joules / kg Kelvin. The sensible heat capacity is calculated through a thermal accumulation model and then pushed into a nonlinear neural network layer for mapping, transforming it into a high-dimensional virtual heat storage capacity.

[0089] The processor independently calculates the sensible heat capacity of each temperature-controlled zone according to the aforementioned heat accumulation model and converts it into a dedicated virtual heat storage capacity. Since the commercial building contains numerous interconnected office areas, the processor abstracts each independent temperature-controlled zone as a vertex in a graph network, and sets the structural thermal conductivity of the partition walls as the connection weight.

[0090] Relying on graph attention network mechanisms, data from each temperature control zone permeates and is weighted and aggregated along the connecting lines to adjacent areas. Through this spatial-level fusion, the program can accurately capture the overall thermal state of the entire building when exposed to afternoon sunlight or sudden increases in local population disturbances.

[0091] The communication gateway periodically obtains dynamic carbon emission factors released by the external power grid and uses a time series prediction model to extract the corresponding carbon emission penalty values.

[0092] The virtual heat storage, global thermal state, and carbon emission penalty value are concatenated to form a composite state vector. This vector, along with the environmental static bias term calculated from the indoor target set temperature, is input into a deep reinforcement learning network based on a near-end policy optimization algorithm. After forward inference and layer-by-layer calculation, the network outputs the initial command for adjusting the opening of the chilled water valve.

[0093] To prevent data loss or extreme high temperatures from causing the model output to go out of bounds, the device will synchronously push the composite state vector into the variational autoencoder.

[0094] By using data reconstruction error to quantify the uncertainty of the current working condition, once an anomaly is detected, the state confidence will be significantly reduced and a physical attenuation coefficient will be generated. This coefficient will impose restrictions on the initial instructions and be proportionally reorganized with the underlying traditional conservative control logic.

[0095] The combined safety control commands are converted into a continuous DC voltage of 0 to 10 volts by a digital-to-analog converter chip. This voltage signal is then sent to the stepper motor driver of the water pipe valve. The driver's built-in V / F (voltage / frequency) conversion circuit converts the voltage signal into a pulse signal of the corresponding frequency, thereby smoothly driving the stepper motor.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A building air conditioning intelligent control method based on deep reinforcement learning, characterized in that, include: Obtain ambient temperature, surface heat transfer coefficient, heat transfer area, and specific heat capacity; The process involves: inputting ambient temperature and surface heat transfer coefficient into the flux model to extract convective heat flux; combining this with heat transfer area and specific heat capacity to perform heat accumulation and extract sensible heat capacity; inputting the sensible heat capacity into the mapping space to generate virtual heat storage; specifically: obtaining the reference temperature, capacity cutoff boundary, wall thickness, material density, and time step; in the flux model, performing a difference operation between the ambient temperature and the reference temperature to generate a temperature deviation value; multiplying the surface heat transfer coefficient with the temperature deviation value to obtain the surface heat flux; extracting convective heat flux through the product operation of surface heat flux and heat transfer area; extracting the structural volume base by multiplying the heat transfer area with the wall thickness; multiplying the structural volume base with the material density to generate the solid mass base; multiplying the solid mass base with the specific heat capacity to obtain the heat capacity reference base; extracting the absolute amount of heat exchange through the product operation of convective heat flux and the time step; and introducing the absolute amount of heat exchange into the heat capacity reference base to perform a quotient operation to generate the temperature change. Convert the temperature change into a normalized heat state; The normalized thermal state is introduced into the capacity cutoff boundary to perform boundary constraints and generate sensible heat capacity. Input the sensible heat capacity into the mapping space to obtain the capacity mapping weight parameters; perform a dot product operation on the sensible heat capacity and the capacity mapping weight parameters to generate the weighted capacity; Perform a nonlinear transformation on the weighted capacity to extract the capacity activation value; Obtain the environmental penalty compensation coefficient; perform an aggregation operation using the capacity activation value and the environmental penalty compensation coefficient to obtain the compensated capacity; The compensated capacity is scaled to produce a scaled capacity. Obtain the upper and lower capacity saturation thresholds; combine the scaled capacity with the upper and lower capacity saturation thresholds to apply extreme value constraints, and obtain the legal interval capacity. Perform a smoothing and denoising operation on the legal interval capacity to generate virtual heat storage; Obtain the building topology and structural thermal conductivity; set the virtual heat storage as node features and the structural thermal conductivity as edge weights, and construct a spatial network based on the building topology; perform neighborhood aggregation of node features along edge weights within the spatial network to extract the global thermal state; specifically: parse the building topology to extract vertex coordinates and connection positions; map the virtual heat storage to vertex coordinates to generate an initial node feature map; assign the structural thermal conductivity to connection positions to generate an initial edge feature map. The initial node feature map and the initial edge feature map are combined to construct the original spatial network; the initial node feature map in the original spatial network is combined with the initial edge feature map to perform neighbor node feature transfer to generate neighborhood aggregation features; the coefficient allocation operation is performed on the neighborhood aggregation features to obtain attention enhancement features; Attention-enhanced features are concatenated with the initial node feature map to generate cascaded fused node features; global pooling is then performed on the cascaded fused node features to extract the global hot state; Obtain the power grid carbon emission factor; input the power grid carbon emission factor into the time series model to extract the carbon emission penalty value, and concatenate it with the virtual heat storage and global thermal state to generate a composite state vector; specifically: obtain the benchmark carbon emission reference index; based on the time series model, perform sequence analysis on the power grid carbon emission factor in combination with the benchmark carbon emission reference index to generate time series carbon emission characteristics; Numerical amplification is performed on the time-series carbon emission features to obtain carbon emission penalty values; dimensionality mapping is performed on the virtual heat storage and global thermal state respectively to extract dimensionality-reduced heat storage features and dimensionality-reduced thermal state features; dimensionality concatenation is performed on the carbon emission penalty values, dimensionality-reduced heat storage features, and dimensionality-reduced thermal state features to generate joint state features; cross-correlation operations are performed on the joint state features to generate a composite state vector; environmental static bias terms are extracted; the composite state vector and environmental static bias terms are input into the decision network to output initial instructions; The composite state vector is input into the variational autoencoder to perform reconstruction and extract the reconstruction error; the reconstruction error is input into the confidence evaluation channel to perform uncertainty quantization to generate the state confidence; the state confidence is introduced into the boundary channel to perform amplitude limiting and clipping to generate the physical attenuation coefficient; the initial command is combined with the physical attenuation coefficient to perform amplitude scaling to generate the safety control command.

2. The intelligent control method for building air conditioning based on deep reinforcement learning according to claim 1, characterized in that, The specific process for generating the initial instruction includes: acquiring historical operating parameters and benchmark design parameters; performing deviation quantization calculations on the historical operating parameters and benchmark design parameters to generate static difference features; performing numerical transformation on the static difference features to obtain environmental static bias terms; inputting the composite state vector combined with the environmental static bias terms into the state input layer within the decision network to extract the action hidden states; using the value evaluation layer to perform expectation estimation on the action hidden states to obtain value evaluation coefficients; combining the value evaluation coefficients to perform distribution sampling on the action hidden states to generate action space distribution; and parsing the action space distribution to output the initial instruction.

3. The intelligent control method for building air conditioning based on deep reinforcement learning according to claim 1, characterized in that, The specific process for generating the state confidence score includes: introducing the composite state vector into the encoding channel of the variational autoencoder to perform distribution prediction, generating mean features and variance features; combining the mean features and variance features to perform sampling operations and extract latent variables in the latent space; performing inverse mapping on the latent variables in the latent space to obtain the reconstructed state vector; performing distance calculation on the reconstructed state vector and the composite state vector to extract the reconstruction error; performing distribution evaluation and estimation on the reconstruction error to generate an uncertainty index; and performing inverse numerical transformation on the uncertainty index to generate the state confidence score.

4. The intelligent control method for building air conditioning based on deep reinforcement learning according to claim 3, characterized in that, The specific generation process of the safety control command includes: obtaining a preset upper confidence threshold, a lower confidence threshold, and a safety baseline command; performing interval verification and filtering on the state confidence combined with the upper and lower confidence thresholds to extract a valid confidence level; performing a scaling transformation on the valid confidence level to generate a physical attenuation coefficient; performing a difference operation between the initial command and the safety baseline command to generate a command bias amplitude value; combining the command bias amplitude value and the physical attenuation coefficient, performing a multiplication reduction operation to generate a safety bias amplitude value; performing a command recombination and superposition operation on the safety baseline command and the safety bias amplitude value to extract the safety control command; and converting the safety control command into a physical control signal and sending it to the underlying mechanism.

Images