Energy-saving control method and system for indoor unit of multi-connected system

Abstract

This invention discloses an energy-saving control method and system for indoor units of a multi-split air conditioning system, relating to the field of air conditioning energy-saving control technology. The method includes inputting environmental situational awareness data into a temporal attention model, extracting temperature change matrices for different regions using convolutional layers, extracting the time dependency matrix of indoor unit airflow using a bidirectional LSTM layer, and fusing these matrices using a multi-head self-attention mechanism. Simultaneously, a Gaussian process regression algorithm is used to generate an airflow prediction demand table. The airflow prediction demand table is combined with a 3D point cloud of the building to construct a dynamic airflow topology map. The PSO algorithm is used to iteratively search the dynamic airflow topology map to obtain air supply parameter combinations. Simultaneously, the air supply parameter combinations are mapped to an execution instruction set through a device protocol mapping library. This invention, by integrating environmental situational awareness data and performing online incremental updates to the temporal attention model, enables the energy-saving control scheme to be dynamically adjusted and optimized based on actual operating results, significantly improving the energy efficiency of air conditioning equipment.

Description

Technical Field

[0001] This invention relates to the field of air conditioning energy-saving control technology, and in particular to an energy-saving control method and system for the indoor unit of a multi-split air conditioning system. Background Technology

[0002] Variable Refrigerant Flow (VRF) systems, as an important component of modern building air conditioning solutions, are characterized by high efficiency and flexibility, and are therefore widely used in commercial and residential buildings. With advancements in technology, especially in artificial intelligence, the Internet of Things, and big data analytics, the control methods for VRF systems are constantly evolving.

[0003] However, existing energy-saving control methods for indoor units in multi-split air conditioning systems still have some shortcomings. On the one hand, the control logic of existing multi-split systems is still based on static setpoints or simple feedback adjustment mechanisms, failing to fully integrate multi-dimensional information such as indoor and outdoor environmental parameters, equipment status, and building structure. On the other hand, existing systems typically adopt an open-loop control method, meaning that control commands are formulated based on the current state but no feedback correction is made based on the actual execution effect. This lacks an online model update mechanism based on real-time power consumption data, making it unable to cope with problems such as equipment aging and sudden environmental changes that occur during long-term operation, thus affecting the overall stability and energy efficiency of the system. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an energy-saving control method for the indoor unit of a multi-split air conditioning system to solve the problems of weak environmental sensing capability and lack of real-time feedback and update mechanism.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an energy-saving control method for the indoor unit of a multi-split air conditioning system, which includes inputting environmental situational awareness data into a time-series attention model, extracting temperature change matrices for different regions using a convolutional layer, extracting the time dependence matrix of indoor unit air volume using a bidirectional LSTM layer, and fusing them using a multi-head self-attention mechanism, while simultaneously using a Gaussian process regression algorithm to generate an air volume prediction demand table.

[0008] By combining the air volume forecast demand table with the building's 3D point cloud, a dynamic airflow topology map is constructed. The PSO algorithm is then used to iteratively search the dynamic airflow topology map to obtain the air supply parameter combination. Simultaneously, the air supply parameter combination is mapped into an execution instruction set through the device protocol mapping library.

[0009] The execution instruction set is sent to the equipment control center, and the variable frequency fan is driven by the PID controller to perform air flow regulation. Real-time equipment power consumption data is collected synchronously and input into the time-series attention model for incremental update to obtain the optimized time-series attention model.

[0010] Based on the optimized temporal attention model, the NSGA-III algorithm is used to solve multiple objectives and generate the final indoor unit energy-saving scheme.

[0011] As a preferred embodiment of the indoor unit energy-saving control method of the multi-split air conditioning system described in this invention, the environmental situation perception data includes indoor and outdoor environmental parameters, equipment operating status parameters, equipment power consumption data, and building three-dimensional point cloud data.

[0012] As a preferred embodiment of the indoor unit energy-saving control method for the multi-split air conditioning system described in this invention, the specific steps for generating the air volume prediction demand table using a Gaussian process regression algorithm are as follows:

[0013] A temporal attention model is constructed by integrating convolutional layers and bidirectional LSTM layers through a layered stacking strategy, and environmental situational awareness data is input into the temporal attention model through the MQTT protocol interface.

[0014] The convolutional layer captures local spatiotemporal features through one-dimensional dilated convolution; it uses residual connections to nonlinearly enhance the local spatiotemporal features and uses Softmax normalization to distribute weights to obtain the temperature change matrix.

[0015] The forward LSTM of the bidirectional LSTM layer performs forward time-series iteration through a gating mechanism to generate a forward state sequence; the backward LSTM uses an inverse gating mechanism to perform reverse time-series iteration to obtain a reverse state sequence; the forward and reverse state sequences are fused using a weighted average method to generate a time dependency matrix.

[0016] Multi-head self-attention mechanism is used to project features onto temperature change matrix and time dependence matrix, and spatiotemporal fusion features are generated by concatenating them through cross-modal channels;

[0017] A Gaussian process regression algorithm is used to perform nonlinear mapping on the spatiotemporal fusion features to generate a wind volume forecast demand table.

[0018] As a preferred embodiment of the indoor unit energy-saving control method for the multi-split air conditioning system described in this invention, the step of constructing a dynamic airflow topology map specifically includes the following steps.

[0019] The air volume forecast demand table is discretized into a grid using a spatial coding algorithm to obtain an air volume distribution matrix. The air volume distribution matrix and the building 3D point cloud are then linked by the A* algorithm to update the node and edge weights, generating a dynamic airflow topology map.

[0020] As a preferred embodiment of the indoor unit energy-saving control method for the multi-split air conditioning system described in this invention, the step of mapping the air supply parameter combination into an execution instruction set through a device protocol mapping library specifically includes the following steps.

[0021] The PSO algorithm is used to perform multidimensional parameter space decomposition on the dynamic airflow topology map to obtain the global particle swarm state; dynamic inertia weight adjustment and neighborhood mutation are used to iteratively search the global particle swarm state to obtain the air supply parameter combination.

[0022] Based on the device protocol mapping library, the air supply parameters are combined and encoded into hexadecimal instructions using Modbus function codes to obtain the original instruction frame; the original instruction frame is then timestamped using a clock synchronization method to generate an execution instruction set.

[0023] As a preferred embodiment of the indoor unit energy-saving control method for the multi-split air conditioning system described in this invention, the step of obtaining the optimized temporal attention model specifically includes the following steps.

[0024] The execution instruction set is sent to the device control center via the CoAP protocol, and the execution instruction set is prioritized using a priority scheduling algorithm to generate a timing instruction queue.

[0025] The PID controller converts the timing command queue into PWM duty pulses through proportional-integral-derivative operations; the PWM duty pulses drive the variable frequency fan to regulate the airflow by adjusting the armature voltage.

[0026] The device power consumption data is collected using a current sensing resistor and input into a timing attention model via an RS485 interface.

[0027] The temporal attention model assigns weights to device power consumption data through Softmax normalization to obtain a power consumption weight vector. Based on the power consumption weight vector, the GRU parameters are incrementally updated using the sliding window incremental learning method to output the optimized temporal attention model.

[0028] As a preferred embodiment of the indoor unit energy-saving control method for the multi-split air conditioning system described in this invention, the step of generating the final indoor unit energy-saving scheme specifically includes the following steps.

[0029] The optimized temporal attention model is dynamically focused by embedding causal convolution to generate a temporal energy distribution matrix.

[0030] The NSGA-III algorithm performs multi-objective optimization of the time-series energy distribution matrix through hierarchical non-dominated sorting to obtain the Pareto solution set; it uses the entropy weight method to normalize the weights of the Pareto solution set to obtain the optimal energy-saving parameters; and it uses the Unreal Engine to perform 3D rendering to generate the final indoor unit energy-saving scheme.

[0031] Secondly, the present invention provides an indoor unit energy-saving control system for a multi-split air conditioning system, comprising an air volume prediction module, an instruction generation module, a model optimization module, and a scheme generation module.

[0032] The air volume prediction module is used to input environmental situational awareness data into the temporal attention model, the convolutional layer extracts the temperature change matrix of different regions, the bidirectional LSTM layer extracts the time dependence matrix of indoor unit air volume, and uses a multi-head self-attention mechanism to fuse them, and simultaneously uses the Gaussian process regression algorithm to generate an air volume prediction demand table.

[0033] The instruction generation module is used to combine the air volume forecast demand table with the building's 3D point cloud to construct a dynamic airflow topology map, and use the PSO algorithm to iteratively search the dynamic airflow topology map to obtain the air supply parameter combination. Simultaneously, the air supply parameter combination is mapped into an execution instruction set through the device protocol mapping library.

[0034] The model optimization module is used to send the execution instruction set to the equipment control center, drive the variable frequency fan to perform air flow regulation through the PID controller, collect real-time equipment power consumption data synchronously, and input it into the time-series attention model for incremental update to obtain the optimized time-series attention model.

[0035] The scheme generation module is used to generate the final indoor unit energy-saving scheme by solving multiple objectives using the NSGA-III algorithm based on the optimized temporal attention model.

[0036] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the indoor unit energy-saving control method of the multi-unit air conditioning system as described in the first aspect of the present invention.

[0037] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the indoor unit energy-saving control method of the multi-unit air conditioning system as described in the first aspect of the present invention.

[0038] The beneficial effects of this invention are as follows: By integrating multi-dimensional information such as indoor and outdoor environmental parameters, equipment status, and building 3D point cloud data, it significantly enhances the perception and response speed to complex indoor environmental changes, ensuring more precise and efficient temperature regulation. Simultaneously, based on real-time power consumption data, the time-series attention model is updated incrementally online, enabling the energy-saving control scheme to be dynamically adjusted and optimized according to actual operating results, greatly improving the stability and energy efficiency of air conditioning equipment. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of the energy-saving control method for the indoor unit of a multi-split air conditioning system.

[0041] Figure 2 This is a schematic diagram of the indoor unit energy-saving control system of a multi-split air conditioning system.

[0042] Figure 3 This is a flowchart of the process for obtaining the air volume forecast demand table.

[0043] Figure 4 A flowchart of the process for constructing and optimizing dynamic airflow topology maps. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0047] Reference Figures 1-4 As an embodiment of the present invention, this embodiment provides an energy-saving control method for the indoor unit of a multi-split air conditioning system, comprising the following steps:

[0048] S1. Input environmental situational awareness data into the temporal attention model, extract the temperature change matrix of different regions through the convolutional layer, extract the time dependence matrix of indoor unit air volume through the bidirectional LSTM layer, and fuse them using a multi-head self-attention mechanism. Simultaneously, use the Gaussian process regression algorithm to generate an air volume prediction demand table.

[0049] Specifically, the steps include the following:

[0050] S1.1 Collect environmental situational awareness data, including indoor and outdoor environmental parameters, equipment operating status parameters, equipment power consumption data, and building 3D point clouds. Indoor and outdoor environmental parameters are collected using multiple sensors deployed inside and outside the building: Indoors, temperature and humidity sensors are installed in key locations such as areas with frequent human activity and air conditioning vents to collect indoor temperature and humidity data; light sensors are installed near windows and on ceilings with natural light sources to collect light data. Outdoors, temperature and humidity sensors are installed on the building's exterior walls to collect outdoor temperature and humidity data; wind speed sensors are installed on the roof to collect wind speed data. The collected temperature and humidity data, light data, and wind speed data are integrated using the MQTT communication protocol to generate indoor and outdoor environmental parameters.

[0051] Equipment operating status parameters are collected through intelligent sensors and controllers integrated into the multi-split air conditioner: the pressure sensor, temperature sensor and frequency sensor built into the indoor unit continuously monitor key operating parameters, including the compressor operating frequency, condenser and evaporator temperatures, and refrigerant pressure; the microprocessor on the control board also records the equipment's on / off status and fault alarm information; the key operating parameters and status parameters are fused using a weighted fusion method to generate equipment operating status parameters;

[0052] The power consumption data of the equipment is collected through a current sensing device, which includes a high-precision energy meter and a current transformer. The current sensing device is installed in the power supply lines of the main components of the multi-split unit (such as the compressor and fan motor). The current sensing device can directly read electrical parameters such as voltage, current and power factor, and use Discrete Fourier Transform (DFT) to calculate the power of the electrical parameters to obtain the total energy consumption value. The energy consumption value is accumulated over time using an integral algorithm to obtain the equipment power consumption data.

[0053] The 3D point cloud of the building is acquired using a laser scanner. The laser scanner is installed on the exterior wall of the building to scan the building structure and its surrounding environment and emit laser beams, while simultaneously receiving the reflected echo signals. TrimbleRealWorks is used to filter, register, and stitch the echo signals to generate a high-precision point cloud. For hard-to-reach areas of the building facade, a drone equipped with a camera is used to perform oblique photography to generate supplementary point clouds. Professional 3D modeling software (such as Autodesk Revit) is used to stitch and merge the high-precision point cloud and the supplementary point cloud to finally form a complete 3D point cloud data of the building.

[0054] S1.2. The collected environmental situation awareness data is preprocessed. Specifically, firstly, mean filtering is used to smooth the environmental situation awareness data with a sliding window and separate nonlinear noise to eliminate sensor noise. Next, K-nearest neighbor interpolation is used to impute missing values ​​in the environmental situation awareness data to obtain a complete spatiotemporal dataset. Subsequently, the complete spatiotemporal dataset is scaled using Min-Max standardization and dimensionality is reduced using PCA. At the same time, timestamp alignment is performed using dynamic time warping (DTW) to obtain the preprocessed environmental situation awareness data.

[0055] S1.3. Constructing and training a temporal attention model. Specifically, in the PyTorch framework, the convolutional layers are first initialized using a two-layer cascaded Conv1D structure: the first convolutional kernel size is set to 5, with 64 channels; the second convolutional kernel size is set to 3, with 128 channels; and batch normalization (BatchNorm1d) and ReLU activation functions are applied after the second convolutional kernel. Next, the bidirectional LSTM layer is initialized using nn.LSTM parameters: the LSTM hidden layer dimension is set to 256, the number of attention heads in the multi-head attention layer is set to 4, and max pooling and Dropout layers are connected after the multi-head attention layer. A layered stacking strategy is used to perform residual connections on the initialized convolutional layers and the bidirectional LSTM layer to obtain high-dimensional feature representations. Finally, fully connected layers are used to integrate the high-dimensional feature representations, completing the construction of the temporal attention model.

[0056] Next, the constructed temporal attention model is trained. Specifically, environmental situational awareness data is input into the temporal attention model via the MQTT protocol interface, and the environmental situational awareness data is divided into a sample set, a training set, and a validation set using a hierarchical random sampling method. On the sample set, a sliding window is used to segment temporal segments to generate standardized training samples. On the training set, the Adam optimizer is used to perform gradient backpropagation on the standardized training samples to obtain the parameter update gradient. The MAE loss function is used to iteratively update the weights of the parameter update gradient to generate the MAE value. On the validation set, when the MAE value reaches the convergence threshold for 10 consecutive epochs, early stopping is triggered and training is terminated. Simultaneously, the torch.save function is used to save the trained temporal attention model.

[0057] It should be noted that the convergence threshold is defined based on the moving average of the MAE value; the value range is [0.001~0.005].

[0058] S1.4. Spatiotemporal fusion features are generated through a temporal attention model. Specifically, the convolutional layer divides the environmental situational awareness data into temporal windows using one-dimensional dilated convolutions and expands the features using causal padding to generate a coarse-grained feature map. Based on the coarse-grained feature map, secondary convolutions are used to abstract features and transform dimensions to obtain local spatiotemporal features. Gated linear units are used to nonlinearly enhance the local spatiotemporal features to obtain a sparse feature matrix. Softmax normalization is used to assign different weights to the sparse feature matrix, and cross-channel stitching is performed through a multi-head attention mechanism to generate a temperature change matrix.

[0059] The forward LSTM of the bidirectional LSTM layer extracts temporal features from environmental situational awareness data through a gating mechanism, and aggregates them through a fully connected layer to generate preliminary hidden states. The Sigmoid activation function is used to assign gating weights to the preliminary hidden states to obtain gating weighted features. The Tanh activation function is used to perform nonlinear transformation on the gating weighted features to generate candidate memory states, and the forget gate and input gate are used for forward temporal iteration to obtain a forward state sequence.

[0060] Backward LSTM uses a backward gating mechanism to extract backward temporal features from environmental situational awareness data and generate backward hidden states; it uses the ReLU activation function to perform a nonlinear transformation on the backward hidden states to obtain the backward hidden states; it uses Layer Normalization to normalize the backward hidden states and generate dynamic feature representations; based on the dynamic feature representations, it uses Gated Linear Units to perform backward temporal iterations and generate backward state sequences.

[0061] A weighted average method is used to assign different weights to the forward and reverse state sequences; based on the different weights, dynamic fusion is performed in the fully connected layer to generate a time dependency matrix;

[0062] Gated feature interaction is used to perform bidirectional interaction between the temperature change matrix and the time dependency matrix, and a multi-head self-attention mechanism is used for multi-granularity weight allocation. Residual connections are used simultaneously for feature enhancement and gradient stabilization to output attention-enhanced features. Linear projection layers are used to project the attention-enhanced features and then concatenate them through cross-modal channels to generate spatiotemporal fusion features.

[0063] S1.5. A Gaussian process regression algorithm is used to perform nonlinear mapping on the spatiotemporal fusion features to generate an air volume forecast demand table. Specifically, the Gaussian process regression algorithm is used to decompose and solve the spatiotemporal fusion features to obtain the predicted mean and variance. Then, Bayesian optimization is used for hyperparameter fitting to obtain a Gaussian prediction distribution. A kernel function is used to smooth the bandwidth of the Gaussian prediction distribution to generate an optimized distribution. Maximum likelihood estimation is used to decompose the probability density of the optimized distribution, and a sliding window filter is used for noise suppression to generate a smooth air volume forecast sequence. Based on the 3σ interval coverage rule, Gaussian kernel regression is used to map the smoothed air volume forecast sequence into an air volume forecast demand table.

[0064] It should be noted that the 3σ interval coverage rule is based on the integral area definition of the normal distribution probability density function;

[0065] The air volume forecast demand table can not only accurately reflect the future trend of air volume demand changes, but also provide early warning of abnormal fluctuations, providing data support for the scheduling of air conditioning equipment.

[0066] S2. Combine the air volume forecast demand table with the building's 3D point cloud to construct a dynamic airflow topology map, and use the PSO algorithm to iteratively search the dynamic airflow topology map to obtain the air supply parameter combination. Simultaneously, map the air supply parameter combination into an execution instruction set through the device protocol mapping library.

[0067] Specifically, the steps include the following:

[0068] S2.1. Combine the air volume forecast demand table with the building's 3D point cloud to construct a dynamic airflow topology map. Specifically, firstly, the air volume forecast demand table is spatially encoded using an octree index using a spatial coding algorithm to generate air volume-related data; then, Kriging interpolation is used to reconstruct the 3D field of the air volume-related data to obtain a continuous air volume field; finally, the finite volume method is used to discretize the continuous air volume field into a grid using flux integration to generate an unstructured grid; and then gradient descent is used to numerically solve the flow field and achieve residual convergence on the unstructured grid to obtain the air volume distribution matrix.

[0069] ICP registration was used to spatially align the airflow distribution matrix with the 3D point cloud of the building, and coordinate system unification was performed using a rigid body transformation matrix to obtain air-structure coupling data. Based on the air-structure coupling data, Delaunay triangulation was used to perform topological connections to obtain the initial airflow network. The A* algorithm was used to search for the optimal path in the initial airflow network to obtain key airflow channels, and graph embedding was used to aggregate node features of the key airflow channels to complete node association and generate a weighted directed graph. The weighted directed graph was then optimized by spectral clustering using the Laplacian matrix to update edge weights and generate a dynamic airflow topology map.

[0070] It should be noted that the rigid body transformation matrix refers to the 4×4 homogeneous coordinate transformation matrix of rotation and translation in three-dimensional space, which is called through the registration_icp function of the Open3D library; the Laplacian matrix refers to the characteristic matrix describing the topological structure in graph theory, which can be calculated and output through the scipy library of the Python engine.

[0071] S2.2. Iteratively search the dynamic airflow topology using the PSO algorithm to obtain the air supply parameter combination. Specifically, the PSO algorithm (Particle Swarm Optimization) is used to decompose the dynamic airflow topology into spectral clustering features to obtain the adjacency matrix features. K-means clustering is used to spatially partition the adjacency matrix features to generate an initial particle distribution. PCA is used to decompose the initial particle distribution into a multidimensional parameter space to generate a dimensionality-reduced search space. In the dimensionality-reduced search space, the initial particle distribution is validated for particle fitness based on a decision threshold. For example, initial particle distributions that exceed the decision threshold are defined as validated and validated initial particle distributions are integrated using an adaptive weighting method to obtain the global particle swarm state.

[0072] It should be noted that the decision threshold particle swarm fitness distribution is defined as the inter-quartile range (IQR), with a value range of [0.7, 0.9].

[0073] Dynamic inertia weight adjustment is used to update the velocity and position of the global particle swarm state, and cross-validation is used for weighted sorting to generate an elite particle set. The Gaussian mutation operator is used to perform neighborhood mutation on the elite particle set to obtain enhanced candidate solutions. Based on the enhanced candidate solutions, Pareto sorting is used for iterative search to generate a non-dominated solution set. TOPSIS decision is used to iteratively optimize the non-dominated solution set to generate the air supply parameter combination.

[0074] It should be noted that the Gaussian mutation operator refers to a probabilistic perturbation operator based on the normal distribution, which performs neighborhood mutation on the elite particle set by randomly sampling N(μ,σ²), with the value range being [μ-3σ, μ+3σ].

[0075] S2.3. The air supply parameter combination is mapped to the execution instruction set through the device protocol mapping library. In specific operation, the device protocol mapping library is called through the OPC UA interface, and the device protocol mapping library is matched with the register address through the function code parser to obtain the Modbus function code; the air supply parameter combination is divided into blocks and packaged according to the Modbus function code to generate structured data; the LZW dictionary encoding method is used to perform repeated string hash replacement on the structured data to generate deduplicated data blocks; the Modbus RTU encoder is used to encode the deduplicated data blocks with hexadecimal instructions to obtain the original instruction frame.

[0076] The original instruction frame is aligned with the time base using the clock synchronization method to generate a time-synchronized instruction frame; the time-synchronized instruction frame is compressed in real time using the LZ4 compression algorithm, and the timestamp is embedded using the NTP protocol. The Modbus TCP protocol is used for encapsulation to generate an execution instruction set.

[0077] The execution instruction set can not only precisely control the speed and airflow direction of the air conditioning equipment, but also intelligently adjust the air supply temperature to quickly bring the indoor temperature to a comfortable level.

[0078] S3. Send the execution instruction set to the equipment control center, drive the variable frequency fan to perform air flow regulation through the PID controller, collect real-time equipment power consumption data synchronously, and input it into the time-series attention model for incremental update to obtain the optimized time-series attention model.

[0079] Specifically, the steps include the following:

[0080] S3.1 The execution instruction set is sent to the equipment control center, and the variable frequency fan is driven by the PID controller to perform air flow regulation. The specific operation is as follows: First, the execution instruction set is encoded into binary TLV through the CoAP protocol to obtain the protocol data unit, and the protocol data unit is sent to the equipment control center in segments using the block transfer option of Block2. At the same time, the retransmission mechanism is used to ensure the reliability of data transmission when the network jitter occurs.

[0081] After receiving the execution instruction set, the control center uses a priority scheduling algorithm to dynamically allocate priorities based on QoS tags, and uses an NTP time server to decompose the execution instruction set into UTC timestamps to obtain the timestamp difference. Based on the timestamp difference, the queue adjuster promotes the queue position of the emergency instructions, and uses the EDF algorithm to perform secondary sorting according to the effective time window of the instructions, finally outputting the time sequence instruction queue.

[0082] It should be noted that the QoS tag refers to the criticality level identifier of the instruction, which is invoked through the protocol parsing unit of the device control center;

[0083] The PID controller is initialized using the Ziegler-Nichols tuning method: the integral and derivative components are set to zero, and the proportional coefficient is gradually increased until the output of the PID controller exhibits constant amplitude oscillation. The critical proportional coefficient and oscillation period at this point are recorded. The output waveform is monitored simultaneously using a step response test. When the output waveform meets the overshoot threshold, the initialization is complete.

[0084] It should be noted that the critical proportional gain refers to the proportional gain value of the PID controller when it produces constant amplitude oscillations under pure proportional control, and is defined based on the closed-loop critical stability of the proportional gain; the overshoot threshold is defined based on the amplitude of the first peak of the step response, and its value ranges from 15% to 20% of the amplitude of the first peak.

[0085] Next, the initialized PID controller is used to compare the setpoint (SP) and process variable (PV) in the timing command queue in real time to obtain the current control deviation. A sliding window filter is then used to suppress noise in the current control deviation, generating a deviation signal value. Finally, proportional-integral-derivative (PI-DE) operations are used to dynamically compensate for the deviation signal value, generating a compensation control quantity. The specific mathematical formula is as follows:

[0086] ;

[0087] in, Indicates a time index. Indicates time The compensation control amount, This represents the proportionality coefficient. Indicates the relationship between setpoints and process variables over time. The filtered difference, Represents the integral coefficient. Indicates the time step. Indicates the relationship between setpoints and process variables at time steps. The filtered difference, Represents the differential coefficient;

[0088] It should be noted that the setpoint refers to the target control quantity of the timing instruction queue, defined based on the actual control task requirements; the process variable refers to the real-time feedback quantity of the timing instruction queue, which is acquired through the signal conditioning circuit; the proportional coefficient is defined based on the actual response speed requirements, with a value range of [0.5-1.2]; the integral coefficient is defined based on the actual steady-state error requirements, with a value range of [0.1-3.0]; and the derivative coefficient is defined based on the actual overshoot suppression requirements, with a value range of [0.01-0.1].

[0089] An output limiter is used to clamp the amplitude of the compensation control quantity to obtain a safe control quantity; a PWM generator is used to perform a duty cycle linear mapping on the safe control quantity to generate a PWM duty pulse.

[0090] The PWM duty pulse is linearly mapped through a digital low-pass filter to obtain the equivalent voltage command; the equivalent voltage command is converted from amplitude to phase using a phase modulator to generate an armature drive signal; based on the armature drive signal, the armature voltage is PWM chopped and modulated using an H-bridge power topology to generate a pulse voltage waveform; the pulse voltage waveform is ripple suppressed and regulated using an LC filter circuit to generate a controllable DC voltage.

[0091] Based on the controllable DC voltage, the speed of the variable frequency fan is modulated by a brushless motor to obtain real-time speed feedback data; based on the real-time speed feedback data, the air supply flow is adjusted by a proportional control algorithm to correct the error and complete the adjustment of the air supply flow.

[0092] S3.2. Synchronously collect real-time device power consumption data and input it into the timing attention model for incremental updates to obtain the optimized timing attention model. The specific operation is as follows: During the air flow adjustment process, the device power consumption data is collected synchronously using the current sensing resistor and input into the timing attention model through the RS485 interface.

[0093] The temporal attention model dynamically weights device power consumption data using Softmax normalization to obtain an attention weight matrix; it randomly samples the attention weight matrix using a sliding window sampling method and expands its dimensions through an embedding layer to obtain a high-dimensional feature vector; it then projects the high-dimensional feature vector onto the matrix using a linear transformation layer to generate a query matrix, a key matrix, and a value matrix; and finally, it assigns dynamic weights to the query matrix, key matrix, and value matrix through a multi-head attention mechanism, simultaneously aggregating the weights to obtain a power consumption weight vector.

[0094] The robustness of the power consumption weight vector is enhanced by adversarial training, and stability is optimized by gradient penalty to obtain anti-interference weights. Based on the obtained anti-interference weights, the GRU parameters are fine-tuned online using sliding window incremental learning to generate dynamic GRU units. The dynamic GRU units are incrementally updated using momentum gradient descent to obtain adaptive network parameters. The adaptive network parameters are weighted and integrated to output the optimized temporal attention model.

[0095] It should be noted that the GRU parameters refer to the update gate weight, reset gate weight, and candidate hidden state weight in the gated recurrent unit, which are obtained through backpropagation over time.

[0096] The optimized temporal attention model can accurately predict the real-time power consumption of air conditioning equipment, laying the foundation for maximizing energy efficiency.

[0097] S4. Based on the optimized temporal attention model, the NSGA-III algorithm is used to solve multiple objectives and generate the final indoor unit energy-saving scheme.

[0098] Specifically, the steps include the following:

[0099] S4.1 The optimized temporal attention model is dynamically focused by embedding causal convolutions to generate a temporal energy distribution matrix. Specifically, firstly, causal convolutions are embedded in the convolutional layers of the optimized temporal attention model through dilation parameters. The causal convolutions are temporally aligned using zero padding to generate causal convolution features. After the causal convolution features are activated by gated linear units (GLUs), they are added to the positional encoding to form a spatiotemporal feature basis.

[0100] It should be noted that positional encoding refers to the sine and cosine function mapping vector of causal convolution features, which is called through the torch.arange function in PyTorch;

[0101] Locality Sensitive Hash (LSH) is used to bucket the spatiotemporal feature basis to reduce data complexity and generate a set of hash buckets with similar features. A sparse attention mechanism is then used to selectively assign weights and associate features with the hash bucket set, achieving dynamic feature focusing and obtaining high-information-density feature representations. Simultaneously, strided convolution is used to locally enhance the high-information-density feature representations, generating multi-scale energy features. Finally, a feature tensor recombination method is used to fuse the spatiotemporal dimensions of the multi-scale energy features to obtain a temporal energy distribution matrix.

[0102] S4.2 The NSGA-III algorithm performs multi-objective optimization on the temporal energy distribution matrix through hierarchical non-dominated sorting to obtain the Pareto solution set. Specifically, the NSGA-III algorithm normalizes the temporal energy distribution matrix and samples reference points to obtain the initial population; the initial population is hierarchically divided into objective spaces through non-dominated sorting to obtain the non-dominated front level; based on the non-dominated front level, a multi-objective elite retention strategy is used to perform multi-objective optimization on the initial population using polynomial mutation and simulated binary crossover to obtain the candidate solution set, and the candidate solution set is niching-assigned to generate a new generation population; the environment selection operator is used to iterate the crowding of the new generation population to obtain the Pareto solution set.

[0103] It should be noted that the environment selection operator is defined based on the target space grid crowding degree, and its value range is [0,1], where 0 indicates that the solutions are completely overlapping and 1 indicates that they are uniformly distributed;

[0104] S4.3. The Pareto solution set is weighted and normalized using the entropy weight method to obtain the optimal energy-saving parameters. Then, 3D rendering is performed using the Unreal Engine to generate the final indoor unit energy-saving solution. Specifically, the Pareto solution set is first divided into energy consumption parameters, comfort parameters, and equipment lifespan index parameters using a sliding window according to multi-objective dimensions. These parameters are then discretized using the entropy weight method to generate a standardized decision matrix. The standardized decision matrix is ​​then sorted and normalized using a weighted summation method to obtain objective weight coefficients. Based on these objective weight coefficients, the Pareto solution set is weighted and standardized to construct the decision matrix. The TOPSIS algorithm is used to measure the bidirectional Euclidean distance between the ideal and negative ideal solutions of the decision matrix to obtain a comprehensive proximity index. Based on the comprehensive proximity index, the decision matrix is ​​validated: when the comprehensive proximity index is greater than the proximity threshold, the decision matrix is ​​considered validated and converted into parameter format using JSON-LD semantics. Simultaneously, the optimal energy-saving parameters are output using the Modbus-TCP protocol.

[0105] It should be noted that the proximity threshold is defined based on the running time of the air conditioning equipment and the real-time operating load rate, and the value range is [0.78, 0.88].

[0106] The optimal energy-saving parameters are imported into the CAD model of the air conditioning equipment using the Datasmith plugin of Unreal Engine, and 3D rendering is performed using Blueprint visualization scripts. Among them, energy consumption parameters are dynamically presented through the self-illumination intensity of materials, comfort parameters drive the color change of virtual PMV sensors, and equipment life index parameters are intuitively displayed through the vibration amplitude and noise particle effects of simulated compressors. Finally, the final indoor unit energy-saving solution demonstration animation is output through Sequencer software, and the effect of parameter optimization can be observed by adjusting the air supply angle and wind speed in real time.

[0107] Ultimately, the indoor unit energy-saving solution can not only effectively reduce the overall energy consumption of air conditioning equipment, but also more accurately regulate the temperature and humidity of the indoor environment, thereby improving the user's comfort experience and extending the service life of the equipment.

[0108] This embodiment also provides an indoor unit energy-saving control system for a multi-split air conditioning system, including: an air volume prediction module, an instruction generation module, a model optimization module, and a scheme generation module.

[0109] The air volume prediction module is used to input environmental situational awareness data into the temporal attention model, the convolutional layer extracts the temperature change matrix of different regions, the bidirectional LSTM layer extracts the time dependence matrix of indoor unit air volume, and uses a multi-head self-attention mechanism to fuse them, and simultaneously uses the Gaussian process regression algorithm to generate an air volume prediction demand table.

[0110] The instruction generation module is used to combine the air volume forecast demand table with the building's 3D point cloud to construct a dynamic airflow topology map, and use the PSO algorithm to iteratively search the dynamic airflow topology map to obtain the air supply parameter combination. Simultaneously, the air supply parameter combination is mapped into an execution instruction set through the device protocol mapping library.

[0111] The model optimization module is used to send the execution instruction set to the equipment control center, drive the variable frequency fan to perform air flow regulation through the PID controller, collect real-time equipment power consumption data synchronously, and input it into the time-series attention model for incremental update to obtain the optimized time-series attention model.

[0112] The scheme generation module is used to generate the final indoor unit energy-saving scheme by solving multiple objectives using the NSGA-III algorithm based on the optimized temporal attention model.

[0113] This embodiment also provides a computer device applicable to the indoor unit energy-saving control method of a multi-split air conditioning system, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the indoor unit energy-saving control method of the multi-split air conditioning system proposed in the above embodiment.

[0114] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0115] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the indoor unit energy-saving control method for a multi-unit air conditioning system as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0116] In summary, this invention significantly enhances the system's ability to perceive and respond to complex indoor environmental changes by integrating multi-dimensional information such as indoor and outdoor environmental parameters, equipment status, and building 3D point cloud data, ensuring more precise and efficient temperature regulation. Simultaneously, online incremental updates to the temporal attention model based on real-time power consumption data enable dynamic adjustments and optimizations to the energy-saving control scheme based on actual operating results, greatly improving the stability and energy efficiency of air conditioning equipment.

[0117] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for energy-saving control of the indoor unit of a multi-split air conditioning system, characterized in that: include, Environmental situational awareness data is input into a temporal attention model. The convolutional layer extracts the temperature change matrix of different regions, the bidirectional LSTM layer extracts the time dependence matrix of indoor unit air volume, and the multi-head self-attention mechanism is used for fusion. At the same time, the Gaussian process regression algorithm is used to generate an air volume prediction demand table. The specific steps for generating the air volume forecast demand table using the Gaussian process regression algorithm are as follows. A temporal attention model is constructed by integrating convolutional layers and bidirectional LSTM layers through a layered stacking strategy, and environmental situational awareness data is input into the temporal attention model through the MQTT protocol interface. The convolutional layer captures local spatiotemporal features through one-dimensional dilated convolution; it uses residual connections to nonlinearly enhance the local spatiotemporal features and uses Softmax normalization to distribute weights to obtain the temperature change matrix. The forward LSTM of the bidirectional LSTM layer performs forward time-series iteration through a gating mechanism to generate a forward state sequence; Backward LSTM uses a reverse gating mechanism to perform reverse time-series iteration and obtain the reverse state sequence; The forward and reverse state sequences are fused using a weighted average method to generate a time dependency matrix; Multi-head self-attention mechanism is used to project features onto temperature change matrix and time dependence matrix, and spatiotemporal fusion features are generated by concatenating them through cross-modal channels; A Gaussian process regression algorithm is used to perform nonlinear mapping on the spatiotemporal fusion features to generate a wind volume forecast demand table. By combining the air volume forecast demand table with the building's 3D point cloud, a dynamic airflow topology map is constructed. The PSO algorithm is then used to iteratively search the dynamic airflow topology map to obtain the air supply parameter combination. Simultaneously, the air supply parameter combination is mapped into an execution instruction set through the device protocol mapping library. The construction of the dynamic airflow topology map specifically includes the following steps. The air volume forecast demand table is discretized into a grid using a spatial coding algorithm to obtain the air volume distribution matrix. The air volume distribution matrix and the building's 3D point cloud are linked via A The algorithm performs node association and edge weight update to generate a dynamic airflow topology graph; The process of mapping the air supply parameter combinations into an execution instruction set through a device protocol mapping library specifically includes the following steps. The PSO algorithm is used to perform multidimensional parameter space decomposition on the dynamic airflow topology map to obtain the global particle swarm state; dynamic inertia weight adjustment and neighborhood mutation are used to iteratively search the global particle swarm state to obtain the air supply parameter combination. Based on the device protocol mapping library, the air supply parameters are combined and encoded into hexadecimal instructions using Modbus function codes to obtain the original instruction frame; the original instruction frame is then timestamped using a clock synchronization method to generate an execution instruction set. The execution instruction set is sent to the equipment control center, and the variable frequency fan is driven by the PID controller to adjust the air supply flow. Real-time equipment power consumption data is collected synchronously and input into the time-series attention model for incremental updates, resulting in an optimized time-series attention model. The specific steps include the following: The execution instruction set is sent to the device control center via the CoAP protocol, and the execution instruction set is prioritized using a priority scheduling algorithm to generate a timing instruction queue. The PID controller converts the timing instruction queue into PWM duty pulses through proportional-integral-derivative operations; The PWM duty pulse drives the variable frequency fan to regulate the airflow by adjusting the armature voltage. The device power consumption data is collected using a current sensing resistor and input into a timing attention model via an RS485 interface. The temporal attention model uses Softmax normalization to assign weights to device power consumption data and obtain a power consumption weight vector. Based on the power consumption weight vector, the GRU parameters are incrementally updated using the sliding window incremental learning method, and the optimized temporal attention model is output. Based on the optimized temporal attention model, the NSGA-III algorithm is used to solve multiple objectives and generate the final indoor unit energy-saving scheme.

2. The indoor unit energy-saving control method of a multi-split air conditioning system as described in claim 1, characterized in that: The environmental situation awareness data includes indoor and outdoor environmental parameters, equipment operating status parameters, equipment power consumption data, and building 3D point clouds.

3. The indoor unit energy-saving control method of a multi-split air conditioning system as described in claim 1, characterized in that: The process of generating the final indoor unit energy-saving solution specifically includes the following steps. The optimized temporal attention model is dynamically focused by embedding causal convolution to generate a temporal energy distribution matrix. The NSGA-III algorithm performs multi-objective optimization of the temporal energy distribution matrix through hierarchical non-dominated sorting to obtain the Pareto solution set; The Pareto solution set is normalized using the entropy weight method to obtain the optimal energy-saving parameters, and then 3D rendering is performed using the Unreal Engine to generate the final indoor unit energy-saving solution.

4. An indoor unit energy-saving control system for a multi-split air conditioning system, based on the indoor unit energy-saving control method for a multi-split air conditioning system according to any one of claims 1 to 3, characterized in that: It includes a wind volume prediction module, an instruction generation module, a model optimization module, and a solution generation module; The air volume prediction module is used to input environmental situational awareness data into the temporal attention model, the convolutional layer extracts the temperature change matrix of different regions, the bidirectional LSTM layer extracts the time dependence matrix of indoor unit air volume, and uses a multi-head self-attention mechanism to fuse them, and simultaneously uses the Gaussian process regression algorithm to generate an air volume prediction demand table. The instruction generation module is used to combine the air volume forecast demand table with the building's 3D point cloud to construct a dynamic airflow topology map, and use the PSO algorithm to iteratively search the dynamic airflow topology map to obtain the air supply parameter combination. Simultaneously, the air supply parameter combination is mapped into an execution instruction set through the device protocol mapping library. The model optimization module is used to send the execution instruction set to the equipment control center, drive the variable frequency fan to perform air flow regulation through the PID controller, collect real-time equipment power consumption data synchronously, and input it into the time-series attention model for incremental update to obtain the optimized time-series attention model. The scheme generation module is used to generate the final indoor unit energy-saving scheme by solving multiple objectives using the NSGA-III algorithm based on the optimized temporal attention model.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the indoor unit energy-saving control method of any one of claims 1 to 3 for a multi-split air conditioning system.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the indoor unit energy-saving control method of any one of claims 1 to 3 for multi-split air conditioning systems.

Images