A method and system for coordinated control of a multi-split air conditioner

Abstract

The embodiment of the application discloses a kind of coordinated control method and system of multiple connected air conditioner, it is applied to outdoor unit, outdoor unit is connected with multiple indoor units, method includes: receiving the time sequence characteristic vector uploaded by each indoor unit, determines global heat load prediction value;According to global heat load prediction value, the recommended refrigerant flow rate and recommended fan speed of each indoor unit are generated;Collect all indoor unit broadcasted local decision broadcast frame, local decision broadcast frame includes the comfort deviation and current fan speed of each indoor unit;According to global heat load prediction value, and the comfort deviation, current fan speed, recommended refrigerant flow rate and recommended fan speed of each indoor unit, in the refrigerant constraint condition is satisfied with integrated energy efficiency optimal as target execution global optimization, generate the recommended control instruction corresponding to the indoor unit to be optimized, and the recommended control instruction is issued to corresponding indoor unit.No need to rely on cloud or independent group controller, communication delay is significantly reduced, and system real-time is improved.

Description

Technical Field

[0001] This invention relates to the field of air conditioning collaborative control technology, and in particular to a collaborative control method for multi-split air conditioners, a collaborative control system for multi-split air conditioners, a multi-split air conditioner, and a computer-readable storage medium. Background Technology

[0002] Currently, multi-split air conditioning systems generally rely on a cloud / group controller or an outdoor unit main control unit to centrally complete global decisions. The outdoor unit, as the main control unit, is equipped with a high-performance DSP (Digital Signal Processor) or ARM (Advanced RISC Machine) chip, responsible for the compressor's variable frequency drive and refrigerant circulation control. The indoor unit, as a slave terminal, is only equipped with a low-power MCU (Microcontroller Unit), responsible for receiving the set temperature, driving the fan, and reporting limited parameters such as return air temperature to the outdoor unit via the bus.

[0003] However, existing multi-split air conditioning control relies on cloud or centralized controllers, which suffer from large communication delays and high risk of single-point failures. Under the cloud-based centralized computing or independent group controller architecture, network links are subject to jitter, packet loss, and bandwidth limitations. Furthermore, failure of centralized nodes will lead to global loss of control, and control command delays can reach the second or even minute level, making it difficult to meet the requirements of high real-time performance and high reliability of intelligent collaborative control. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention are proposed to provide a collaborative control method for a multi-split air conditioner, a collaborative control system for a multi-split air conditioner, a multi-split air conditioner, and a computer-readable storage medium to overcome or at least partially solve the above problems.

[0005] To address the aforementioned problems, a first aspect of this invention provides a collaborative control method for a multi-split air conditioner, applied to an outdoor unit connected to multiple indoor units, the method comprising: Receive the time-series feature vectors uploaded by each indoor unit, and determine the global heat load prediction value based on the time-series feature vectors of each indoor unit; Based on the global heat load forecast, a recommended refrigerant flow rate and recommended fan speed are generated for each indoor unit; Collect local decision broadcast frames from all indoor units, wherein each local decision broadcast frame contains the comfort deviation and current fan speed of each indoor unit; Based on the global heat load forecast, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate level, and the recommended fan speed, global optimization is performed with the goal of optimal overall energy efficiency under the condition of satisfying refrigerant constraints. Recommended control commands are generated for the indoor units to be optimized and sent to the corresponding indoor units.

[0006] According to a second aspect of the present invention, a collaborative control method for a multi-split air conditioner is provided, applied to an indoor unit, wherein the indoor unit is connected to an outdoor unit and other indoor units respectively, the method comprising: The collected environmental data is used to extract features to obtain a time-series feature vector, and the time-series feature vector is uploaded to the outdoor unit; The comfort deviation is calculated based on the environmental data and the temporal feature vector. Obtain the current fan speed, and encapsulate the comfort deviation and the current fan speed into a local decision broadcast frame for broadcasting; Receive the recommended control command issued by the outdoor unit; The system outputs local control commands based on the current operating status of the indoor unit, arbitrates the recommended control commands with the local control commands, generates target control commands based on the arbitration result, and executes the target control commands.

[0007] According to a third aspect of the present invention, a coordinated control system for a multi-split air conditioner is provided, the system comprising an outdoor unit and multiple indoor units, wherein the indoor units are respectively connected to the outdoor unit and other indoor units; The indoor unit is used to extract features from the collected environmental data to obtain a time-series feature vector, and then upload the time-series feature vector to the outdoor unit. The outdoor unit is used to receive the time-series feature vectors uploaded by each indoor unit and determine the global heat load prediction value based on the time-series feature vectors of each indoor unit. The indoor unit is used to calculate the comfort deviation based on the environmental data and the time-series feature vector; obtain the current fan speed; and encapsulate the comfort deviation and the current fan speed into a local decision broadcast frame for broadcasting. The outdoor unit is used to generate recommended refrigerant flow rates and recommended fan speeds for each indoor unit based on the global heat load forecast; collect local decision broadcast frames broadcast by all indoor units; and perform global optimization with the goal of optimal overall energy efficiency under the condition of satisfying refrigerant constraints, based on the global heat load forecast, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate, and the recommended fan speed, generating recommended control commands for the indoor units to be optimized, and sending the recommended control commands to the corresponding indoor units. The indoor unit is used to receive recommended control commands from the outdoor unit, output local control commands according to the current operating status of the indoor unit, arbitrate the recommended control commands and the local control commands, generate target control commands based on the arbitration result, and execute the target control commands.

[0008] According to a fourth aspect of the present invention, a multi-split air conditioner is provided, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the cooperative control method for the multi-split air conditioner as described in any of the preceding embodiments.

[0009] According to a fifth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, wherein when executed by a processor, the computer program implements the steps of the cooperative control method for a multi-split air conditioner as described in any of the preceding claims.

[0010] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a collaborative control method and system for a multi-split air conditioner, applied to an outdoor unit connected to multiple indoor units. The method includes: receiving time-series feature vectors uploaded by each indoor unit and determining a global heat load prediction value based on the time-series feature vectors of each indoor unit; generating recommended refrigerant flow rates and recommended fan speeds for each indoor unit based on the global heat load prediction value; collecting local decision broadcast frames broadcast by all indoor units, where each local decision broadcast frame contains the comfort deviation and current fan speed of each indoor unit; and performing global optimization with the goal of optimal overall energy efficiency under refrigerant constraints, based on the global heat load prediction value, the comfort deviation, current fan speed, recommended refrigerant flow rate, and recommended fan speed of each indoor unit, generating recommended control commands for the indoor units to be optimized, and sending the recommended control commands to the corresponding indoor units. The outdoor unit collects local decision broadcast frames from each indoor unit. These frames contain the comfort deviation and current fan speed of each indoor unit. Based on this information, the outdoor unit performs global optimization locally, generates recommended control commands for the indoor units to be optimized, and issues them directly. This eliminates the need to rely on the cloud or a separate group controller, significantly reducing communication latency and improving system real-time performance. By fusing multi-source data from global heat load prediction values ​​and local decision broadcast frames, effective coordination between global optimization and local states is achieved, solving the problems of high communication latency and high single-point failure risk in existing centralized architectures. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the steps of a collaborative control method for a multi-split air conditioner provided in an embodiment of the present invention; Figure 2This is a flowchart of another collaborative control method for multi-split air conditioners provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the computational power unloading decision-making process for a collaborative control method for a multi-split air conditioner provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating a collaborative control method for a multi-split air conditioner provided in an embodiment of the present invention; Figure 5 This is a structural block diagram of a collaborative control system for a multi-split air conditioner provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the above-mentioned objects, features and advantages of the present invention more readily understood, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] Existing multi-split air conditioning control relies on cloud or centralized controllers, which suffer from large communication delays and high risk of single-point failures. Under the cloud-based centralized computing or independent group controller architecture, network links are subject to jitter, packet loss, and bandwidth limitations. Furthermore, failure of centralized nodes will lead to global loss of control, and control command delays can reach the second or even minute level, making it difficult to meet the requirements of high real-time performance and high reliability of intelligent collaborative control.

[0014] One of the core concepts of this invention is that the outdoor unit collects local decision broadcast frames broadcast by each indoor unit. The local decision broadcast frames contain the comfort deviation of each indoor unit and the current fan speed. Based on this, the outdoor unit completes global optimization locally, generates recommended control commands for the indoor unit to be optimized, and sends them directly without relying on the cloud or an independent group controller. This significantly reduces communication latency and improves system real-time performance.

[0015] Reference Figure 1 This diagram illustrates a flowchart of a collaborative control method for a multi-split air conditioner according to an embodiment of the present invention. The method is applied to an outdoor unit connected to multiple indoor units, and specifically includes the following steps: Step 101: Receive the time-series feature vectors uploaded by each indoor unit, and determine the global heat load prediction value based on the time-series feature vectors of each indoor unit. The outdoor unit is connected to multiple indoor units via a main refrigerant pipe and branch pipes, forming a refrigerant circulation loop. The outdoor unit side is equipped with a condenser, inverter compressor, and multi-way electronic expansion valves. Each electronic expansion valve corresponds to one indoor unit, used to independently control the refrigerant flow to each indoor unit. The indoor unit side is equipped with an evaporator and a fan. The refrigerant absorbs heat and evaporates in the evaporator of the indoor unit, achieving cooling (the flow direction is reversed during heating). The refrigerant flow to each indoor unit is independently adjusted by its corresponding electronic expansion valve, thus achieving precise zone control.

[0016] The outdoor unit is connected to each indoor unit via an indoor-outdoor unit collaborative communication bus. Each indoor unit uploads timing feature vectors, local decision broadcast frames, and computing power heartbeat frames to the outdoor unit through this bus; the outdoor unit sends recommended refrigerant flow levels, recommended fan speeds, and control commands to each indoor unit through this bus.

[0017] This embodiment of the multi-split air conditioning system consists of one outdoor unit and N indoor units. The outdoor and indoor units are connected via an enhanced CAN (Controller Area Network) bus and refrigerant piping. The outdoor unit is equipped with an ARM / DSP (Digital Signal Processor) + NPU (Neural Processing Unit) main control unit, running a global scheduling model responsible for global heat load prediction, refrigerant allocation, and computing power scheduling. The indoor unit group is equipped with an MCU (Microcontroller Unit) + NPU edge unit, running a local perception model responsible for local environmental perception, first-level inference, and comfort deviation calculation. In addition to traditional control frames, collaborative computing information such as computing power status frames, task unloading frames, and model synchronization frames are added between the indoor and outdoor units, enabling the communication bus to not only transmit control commands but also carry collaborative computing information.

[0018] The outdoor unit's main control unit is responsible for running the global load scheduling model and compressor frequency optimization algorithm. It connects to multiple electronic expansion valves via the refrigerant main pipe, with each electronic expansion valve corresponding to one indoor unit, enabling independent and precise control of refrigerant flow. The outdoor unit also integrates outdoor ambient temperature sensors, condensing pressure sensors, and exhaust temperature sensors for refrigerant system status monitoring. Each indoor unit is equipped with a 32-bit MCU and an ultra-low-power NPU for running local thermal environment sensing and user behavior recognition models. The indoor unit's sensor array includes a return air temperature sensor, humidity sensor, millimeter-wave human presence sensor, CO2 sensor, and ambient light sensor, all connected to the indoor unit's MCU. Enhanced CAN bus communication is used between the indoor and outdoor units. The high-order bits of the frame ID retain traditional function codes, while the low-order bits add computing power collaboration function codes, ensuring backward compatibility with existing equipment. Through this architecture, the indoor unit is no longer a pure execution terminal but an autonomous decision-making node with local sensing, local reasoning, and pre-decision-making capabilities. The outdoor unit transforms from a single decision-making center into a global scheduling center and computing power scheduling center. The two form a decentralized peer-to-peer collaborative network through model sharding, computing power collaboration frames, and arbitration mechanisms.

[0019] In this embodiment of the invention, each indoor unit is deployed with a first-layer segment of the air conditioning heat load prediction model, which includes an input layer to a hidden layer. The indoor unit inputs environmental data collected by the sensor array into the first-layer segment, extracts time-series feature vectors from the first-layer segment, and uploads them to the outdoor unit via the indoor-outdoor unit collaborative communication bus. The outdoor unit is deployed with a second-layer segment of the same air conditioning heat load prediction model, which includes a hidden layer to an output layer. After receiving the time-series feature vectors uploaded by each indoor unit, the outdoor unit aggregates the time-series feature vectors from each indoor unit, inputs them into the second-layer segment, and outputs the global heat load prediction value. Through the above model segment deployment method, the inference process of the air conditioning heat load prediction model is distributed among the indoor and outdoor units to be completed collaboratively: the indoor units undertake the first-layer feature extraction, and the outdoor units undertake the second-layer global prediction, which utilizes the idle computing power of the indoor units and reduces the computing power burden on the outdoor units.

[0020] The first layer segmentation and the second layer segmentation are two parts of the air conditioning heat load prediction model after being divided into network layers, corresponding to the feature extraction network and the load prediction network, respectively.

[0021] In some embodiments, the outdoor unit and the indoor unit are respectively deployed with pre-trained air conditioning heat load prediction models, the air conditioning heat load prediction models including feature extraction networks and load prediction networks; the time-series feature vectors are extracted by the indoor unit from the original sensor data based on the feature extraction network of the air conditioning heat load prediction model deployed on itself; step 101 may include the following steps: In sub-step S11, the time-series feature vectors of each indoor unit are aggregated and input into the load prediction network, and the global heat load prediction value is output through the load prediction network.

[0022] Air conditioning heat load forecasting models predict the air conditioning heat load demand of individual rooms or the entire building area over a future period based on current and historical indoor and outdoor environmental data, occupant status, and equipment operating status. The predicted heat load serves as the basis for refrigerant flow allocation, compressor frequency adjustment, and power scheduling. The more accurate the forecast, the more proactive the air conditioning system can be in responding to load changes, avoiding delayed adjustments that only become apparent after overcooling / overheating has occurred, thus achieving on-demand cooling and energy-saving operation.

[0023] In this embodiment of the invention, both the outdoor and indoor units are equipped with pre-trained air conditioning heat load prediction models, which include a feature extraction network and a load prediction network. The feature extraction network is deployed on the indoor unit side to extract time-series feature vectors from raw sensor data; the load prediction network is deployed on the outdoor unit side to output a global heat load prediction value based on the time-series feature vectors of each indoor unit. Specifically, the indoor unit collects environmental data such as temperature, humidity, human presence, CO2, and light intensity through a sensor array, inputs this data into its own deployed feature extraction network, which outputs a time-series feature vector and uploads it to the outdoor unit via the indoor-outdoor unit collaborative communication bus.

[0024] After receiving the time-series feature vectors uploaded by each indoor unit, the outdoor unit aggregates these vectors. This involves merging the time-series feature vectors extracted from multiple indoor units at the same time or within the same period into a unified feature representation. The aggregated feature is then input into its deployed load forecasting network. The load forecasting network performs global inference on the aggregated multi-indoor unit features, outputting a global heat load forecast value. This forecast value represents the total heat load demand of the building area served by the entire multi-split air conditioning system over a future period, serving as the basis for subsequently generating recommended refrigerant flow rates and fan speeds for each indoor unit, as well as for performing global optimization. Through the cross-node division of labor between the feature extraction network and the load forecasting network, the inference process of the air conditioning heat load forecasting model is distributed among the indoor and outdoor units, utilizing the idle computing power of the indoor units while reducing the computing burden on the outdoor units.

[0025] Step 102: Generate recommended refrigerant flow rate and recommended fan speed for each indoor unit based on the global heat load prediction value; In this embodiment of the invention, the global heat load forecast represents the total heat load demand of the building area served by the entire multi-split air conditioning system over a future period. After obtaining the global heat load forecast, the outdoor unit needs to break it down to each indoor unit to determine how much refrigerant flow each indoor unit should handle and at what speed the fan should operate.

[0026] The outdoor unit determines the refrigerant flow rate level for each indoor unit based on the global heat load forecast and the room characteristics (such as room area, orientation, and historical load percentage) or an average distribution strategy. This level is a discrete or semi-continuous value indicating the relative amount of refrigerant flow the corresponding indoor unit should receive, such as high, medium, and low levels, or several proportionally divided levels. The recommended refrigerant flow rate level serves as the initial reference value for generating the target opening of the electronic expansion valve through global optimization. The outdoor unit then determines the recommended fan speed for each indoor unit based on the global heat load forecast and the room characteristics and comfort requirements of each indoor unit. This speed indicates the appropriate operating speed for the corresponding indoor unit's fan to coordinate with the refrigerant flow rate level for effective heat exchange in the room. The recommended fan speed also serves as the initial reference value for global optimization.

[0027] In some embodiments, step 102 may include the following steps: Sub-step S21: Determine the total refrigerant demand and total air volume demand based on the global heat load forecast. Sub-step S22: Obtain the rated capacity or historical load percentage of each indoor unit, and allocate recommended refrigerant flow to each indoor unit according to the rated capacity or historical load percentage of each indoor unit; the sum of the recommended refrigerant amounts of all indoor units is the total refrigerant requirement. Sub-step S23: Determine the fan speed of each indoor unit based on the total air volume requirement and the rated air volume of each indoor unit.

[0028] In this embodiment of the invention, the outdoor unit determines the total refrigerant and total air volume requirements of the entire multi-split air conditioning system over a future period based on the obtained global heat load forecast. The global heat load forecast represents the total heat load of the building area, which the outdoor unit converts into the total refrigerant flow rate (i.e., total refrigerant demand) required by the refrigerant circulation system and the total air volume (i.e., total air volume demand) required by the indoor unit fans. The outdoor unit obtains the rated capacity or historical load percentage of each indoor unit and allocates the total refrigerant demand to each indoor unit according to their rated capacity or historical load percentage, obtaining the recommended refrigerant flow rate for each indoor unit. The sum of the recommended refrigerant flow rates of all indoor units equals the total refrigerant demand. The rated capacity reflects the designed heat exchange capacity of the indoor unit, and the historical load percentage reflects the actual load sharing ratio of the indoor unit in past operation. Allocation based on rated capacity is suitable for scenarios where the system is in initial operation or the load distribution is uniform, while allocation based on historical load percentage is suitable for scenarios where the load distribution is uneven or where sufficient operating data has been accumulated.

[0029] The outdoor unit determines the fan speed of each indoor unit based on the total air volume demand and the rated air volume of each indoor unit. Specifically, the outdoor unit allocates the total air volume demand according to the rated air volume of each indoor unit, obtaining the air volume that each indoor unit should handle. Then, based on the correspondence between indoor unit air volume and fan speed, it determines the recommended fan speed for each indoor unit. The outdoor unit first determines the total refrigerant demand and total air volume demand based on the global heat load forecast. Then, it allocates the refrigerant flow according to the rated capacity or historical load percentage of each indoor unit and determines the fan speed according to the rated air volume. The global forecast value is decomposed layer by layer into the specific control parameters of each indoor unit, eliminating the need for complex iterative calculations and resulting in a fast response speed.

[0030] Step 103: Collect all local decision broadcast frames broadcast by indoor units, wherein the local decision broadcast frames contain the comfort deviation of each indoor unit and the current fan speed; In this embodiment of the invention, each indoor unit, during operation, independently calculates the comfort deviation of its room and obtains the current fan speed based on environmental data collected by its own sensor array and the time-series feature vector output by the first-layer segmentation. The indoor unit encapsulates the comfort deviation and the current fan speed into a local decision broadcast frame, which is broadcast on the indoor-outdoor unit collaborative communication bus at a broadcast period. The outdoor unit collects the local decision broadcast frames broadcast by all indoor units at a global optimization period, thereby obtaining the comfort deviation of each room and the current fan speed of each indoor unit.

[0031] Comfort deviation reflects the current thermal comfort state of each room and serves as the basis for global optimization to determine which rooms need adjustment and in what direction. Comfort deviation is a scalar value; negative values ​​indicate excessive cold, positive values ​​indicate excessive heat, and larger absolute values ​​indicate greater deviation from the comfort zone. Current fan speed reflects the current operating status of each indoor unit and serves as a reference for global optimization calculations of refrigerant flow redistribution and fan speed adjustment. The outdoor unit, by comparing the current fan speed with the recommended fan speed, can determine the gap between the actual operating status of each indoor unit and the target status. Local decision broadcast frames are the carriers of local decision information fed back from the indoor units to the outdoor unit. By collecting local decision broadcast frames, the outdoor unit can obtain the comfort status of each room and the operating status of each indoor unit, serving as local state input for subsequent global optimization. Compared to existing technologies where indoor units only report limited parameters such as return air temperature, the local decision broadcast frames broadcast by the indoor units in this embodiment contain richer local decision information such as comfort deviation and current fan speed, enabling the outdoor unit to consider both the comfort needs of each room and the operating status of each indoor unit in global optimization, achieving more refined zoning control.

[0032] Step 104: Based on the global heat load prediction value, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate level and the recommended fan speed, global optimization is performed with the goal of optimal comprehensive energy efficiency under the condition of satisfying refrigerant constraints. Recommended control commands are generated for the indoor units to be optimized, and the recommended control commands are sent to the corresponding indoor units.

[0033] In this embodiment of the invention, the outdoor unit performs global optimization based on the obtained global heat load prediction value, combined with the collected comfort deviations and current fan speeds of each indoor unit, as well as the generated recommended refrigerant flow rate level and recommended fan speed, with the goal of achieving optimal overall energy efficiency under the condition of satisfying refrigerant constraints. It generates recommended control commands corresponding to the indoor units to be optimized and sends the recommended control commands to the corresponding indoor units.

[0034] The outdoor unit collects local decision broadcast frames from each indoor unit during the global optimization cycle, and then filters out the indoor units that need optimization in this round. These are indoor units whose comfort deviation exceeds a preset deviation range, or whose local computing power margin deviates from the average load rate by a preset threshold. For the indoor units to be optimized, the outdoor unit uses their comfort deviation, current fan speed, recommended refrigerant flow rate, and recommended fan speed, along with the global heat load forecast, as inputs for global optimization. Under the condition of satisfying refrigerant constraints, global optimization is performed with the goal of achieving optimal overall energy efficiency. The refrigerant constraints include the compressor frequency being within the upper and lower limits, the sum of the openings of each electronic expansion valve satisfying the conservation of total refrigerant flow, and the evaporator outlet superheat not being lower than a preset value. Global optimization also uses the comfort deviation of each indoor unit being within a preset deviation range and the difference in load rate between each indoor unit being lower than a preset threshold as control constraints.

[0035] Global optimization generates recommended control commands for the indoor units to be optimized. These commands include the target opening degree of the electronic expansion valve, the target frequency of the compressor, and computing power scheduling commands. The target opening degree of the electronic expansion valve is used to adjust the refrigerant flow of the corresponding indoor unit, the target frequency of the compressor is used to adjust the total refrigerant circulation in the system, and the computing power scheduling commands are used to migrate some inference tasks from indoor units with limited computing power to those with surplus computing power, or to adjust the inference frequency of the corresponding indoor units. The outdoor unit sends the recommended control commands to the corresponding indoor units to be optimized; non-optimized indoor units maintain their current control commands, thus ensuring optimal global energy efficiency while avoiding unnecessary adjustments to normally operating indoor units, reducing communication overhead and frequent equipment operations.

[0036] In some embodiments, the local decision broadcast frame further includes the local computing power margin of the corresponding indoor unit; step 104 may include the following steps: Sub-step S31: Determine whether the comfort deviation of each indoor unit is within the preset deviation range, and identify indoor units whose comfort deviation is not within the preset deviation range as indoor units to be optimized; and / or: Perform a computing power load balancing check based on the local computing power margin of each indoor unit, and identify indoor units whose load rate deviates from the average load rate by more than the preset load rate deviation threshold as indoor units to be optimized. Sub-step S32: Based on the global heat load prediction value, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate level and the recommended fan speed, global optimization is performed with the goal of optimal comprehensive energy efficiency under the condition of satisfying refrigerant constraints. Recommended control instructions are generated for the indoor unit to be optimized and sent to the corresponding indoor unit to be optimized.

[0037] Local computing power margin refers to the unused portion of the indoor unit's local computing resources under its current operating state, representing how many additional inference tasks the indoor unit can handle. The local computing power margin is calculated by the indoor unit's local computing power manager. The local computing power manager periodically evaluates the MCU load rate and NPU load rate, combines this with available memory, calculates the local computing power margin, and encapsulates it into a local decision broadcast frame, which is then broadcast to the indoor / outdoor unit collaborative communication bus along with the comfort deviation and the current fan speed.

[0038] In this embodiment of the invention, the outdoor unit obtains the comfort deviation and local computing power margin from the local decision broadcast frames of each indoor unit, and uses this information to filter out the indoor units that need optimization in this round. There are two filtering methods, which can be used selectively or simultaneously: First, determine whether the comfort deviation of each indoor unit is within a preset deviation range. Indoor units whose comfort deviation is not within the preset deviation range are identified as indoor units to be optimized, meaning rooms with excessively large comfort deviations from the comfort zone require adjustment in this round. Second, perform a computing power load balancing check based on the local computing power margin of each indoor unit. Indoor units whose load rate deviates from the average load rate by more than a preset load rate deviation threshold are identified as indoor units to be optimized, meaning nodes with unbalanced computing power loads require adjustment in this round. Through the above filtering, the outdoor unit identifies the truly needing optimization from all indoor units, avoiding unnecessary adjustments to normally operating non-optimizable indoor units.

[0039] The outdoor unit performs global optimization based on the overall heat load forecast, comfort deviations of each indoor unit, current fan speed, recommended refrigerant flow rate, and recommended fan speed, aiming for optimal overall energy efficiency while meeting refrigerant constraints. It generates recommended control commands for the indoor units to be optimized and sends these commands to them. Non-optimized indoor units retain their current control commands. Global optimization only fine-tunes the indoor units to be optimized, ensuring optimal overall energy efficiency while reducing unnecessary adjustments and communication overhead for normal indoor units.

[0040] In some embodiments, the outdoor unit is equipped with a global scheduling model; the refrigerant constraints include compressor frequency being within upper and lower limits, total refrigerant flow being conserved, evaporator outlet superheat not being lower than a preset value, and computing load balancing; step S32 may include the following steps: Sub-step S321: Construct a state vector, which includes the load rate, comfort deviation, and current fan speed of the indoor unit to be optimized, as well as the compressor frequency, condensing pressure, outdoor temperature, and current electricity price of the outdoor unit; In sub-step S322, the state vector, the global heat load prediction value, the recommended refrigerant flow rate level, and the recommended fan speed are input into the global scheduling model. Under the condition of satisfying the refrigerant constraint, with the goal of optimal comprehensive energy efficiency, the global scheduling model outputs parameter adjustment instructions and computing power scheduling instructions for the indoor unit to be optimized. The parameter adjustment instructions are used to instruct the opening degree of the electronic expansion valve of the indoor unit to be optimized to be set to the target opening degree and the compressor frequency to be set to the target frequency. The computing power scheduling instructions are used to migrate some inference tasks of the indoor units with load rates higher than the preset load rate threshold to the indoor units with load rates lower than the preset load rate threshold, or adjust the inference frequency of the corresponding indoor units.

[0041] In this embodiment of the invention, the outdoor unit is equipped with a global scheduling model for performing global optimization. Refrigerant constraints include compressor frequency within upper and lower limits, total refrigerant flow conservation, evaporator outlet superheat not lower than a preset value, and balanced computing load.

[0042] The outdoor unit constructs a state vector, which includes the load rate, comfort deviation, and current fan speed of the indoor unit to be optimized, as well as the outdoor unit compressor frequency, condensing pressure, outdoor temperature, and current electricity price. The load rate, comfort deviation, and current fan speed of the indoor unit to be optimized are derived from local decision broadcast frames broadcast by each indoor unit, reflecting the current operating state of the indoor unit to be optimized. The outdoor unit compressor frequency and condensing pressure are derived from the outdoor unit's own inverter driver and pressure sensor, reflecting the current operating state of the refrigerant system. The outdoor temperature is derived from an outdoor ambient temperature sensor, and the current electricity price is derived from an external electricity price signal. Both are incorporated into the state vector as environmental and economic factors, enabling global optimization to take into account environmental changes and electricity price fluctuations. The state vector is a complete description of the system's current operating state and serves as the input basis for the global scheduling model's optimization.

[0043] The outdoor unit inputs its state vector, global heat load forecast, recommended refrigerant flow rate, and recommended fan speed into the global scheduling model. Under refrigerant constraints, and with the goal of optimal overall energy efficiency, the global scheduling model outputs parameter adjustment instructions and computing power scheduling instructions for the indoor unit to be optimized. The global heat load forecast, output by the second-layer segmentation (load forecasting network), serves as the basis for the total refrigerant allocation. The recommended refrigerant flow rate and recommended fan speed, generated from the global heat load forecast, are initial reference values ​​for global optimization, helping to reduce the search space and computational load of the global scheduling model. Under refrigerant constraints, the global scheduling model optimizes for optimal overall energy efficiency, outputting parameter adjustment instructions and computing power scheduling instructions for the indoor unit to be optimized. The parameter adjustment command instructs the electronic expansion valve opening of the indoor unit to be optimized to the target opening and the compressor frequency to the target frequency, thereby adjusting the refrigerant flow of the corresponding indoor unit and the total refrigerant circulation in the system. The computing power scheduling command is used to migrate some inference tasks of indoor units with load rates higher than the preset load rate threshold to indoor units with load rates lower than the preset load rate threshold, or to adjust the inference frequency of the corresponding indoor units, thereby achieving computing power load balancing among the indoor units to be optimized. With the support of the global scheduling model, the outdoor unit incorporates the global heat load forecast, the local state of the indoor unit to be optimized, the recommended values, and environmental and economic factors into the optimization, generating parameter adjustment commands and computing power scheduling commands for the indoor unit to be optimized, achieving optimal overall energy efficiency while ensuring the safety constraints of the refrigerant system and the computing power load balancing.

[0044] In some embodiments, step S322 may include the following steps: Sub-step S3221: Using the global scheduling model with the comprehensive energy efficiency index as the reward target, synchronously output computing power scheduling actions and cooling control actions; the comprehensive energy efficiency index is the ratio of cooling capacity to comprehensive power consumption, and the comprehensive power consumption includes compressor power consumption, the operating power consumption of all chips in the indoor and outdoor units, and the communication bus power consumption; the computing power scheduling actions include executing inference tasks locally, offloading inference tasks to the outdoor unit or other indoor units, and adjusting the frequency of inference tasks executed by the indoor units; the cooling control actions include compressor frequency and electronic expansion valve opening.

[0045] DDPG (Deep Deterministic Policy Gradient) is a reinforcement learning algorithm specifically designed to solve optimization problems in continuous action spaces. In this embodiment, it is used to train a "computing power scheduling agent," which is the decision-making model in the outdoor unit responsible for simultaneously determining computing power scheduling actions and cooling control actions. Actions in air conditioning control (compressor frequency, EEV opening degree, inference frequency, etc.) are continuous values, not discrete "on / off" or "high / medium / low." Traditional reinforcement learning algorithms (such as Q-Learning) can only handle discrete actions and cannot directly output continuous values. DDPG can directly output continuous actions.

[0046] DDPG employs an Actor-Critic architecture. The Actor network acts as the policy network, determining "what actions to take," while the Critic network acts as the value network, evaluating "whether this action is good or bad." The Actor network inputs a 16-dimensional state vector (MCU load rate, NPU load rate, CDI value for each indoor unit, compressor frequency, condensing pressure, outdoor temperature, and current electricity price for the outdoor unit); the output is a 12-dimensional action vector (inference frequency setpoint for each indoor unit, task execution location code, EEV opening adjustment, and compressor frequency adjustment). Its function is to synchronously output computing power scheduling actions and cooling control actions based on the current system state. The Critic network inputs a 16-dimensional state vector and a 12-dimensional action vector; the output is a scalar value score; its function is to evaluate the long-term overall energy efficiency benefits of the actions output by the Actor network, guiding Actor network updates.

[0047] In this embodiment of the invention, the outdoor unit employs the DDPG reinforcement learning algorithm to train a computing power scheduling agent. After pre-training in a digital twin environment, the agent is deployed to run on the outdoor unit's NPU. When the environmental state is input into the agent, the Actor network synchronously outputs two types of actions: computing power scheduling actions and cooling control actions. These two types of actions respectively affect the chip communication power consumption and compressor power consumption, jointly determining the system's overall energy efficiency. The overall energy efficiency and state parameter feedback form a new state, and the Critic network evaluates the joint value.

[0048] The overall system energy efficiency index is defined as the ratio of cooling capacity to total power consumption, i.e., SCOP_total = Q_cooling / (P_compressor + P_chips + P_comm). Here, Q_cooling is the total system cooling capacity, which can be estimated from the difference between the outdoor unit's condensing and evaporating pressures; P_compressor is the compressor power consumption, which can be estimated from the inverter driver's current sampling; P_chips is the operating power consumption of all chips (MCU, NPU, and communication transceiver) in both indoor and outdoor units, which can be estimated through the accumulation of current feedback from the chip power management unit; and P_comm is the communication bus power consumption, which is proportional to the bus occupancy rate and frame length. This overall energy efficiency index simultaneously considers cooling benefits, compressor power consumption, computing chip power consumption, and communication power consumption, enabling the global scheduling model to pursue cooling efficiency while also taking into account the energy cost of computing power collaboration itself.

[0049] The state vector s can be designed to be 16-dimensional. For example, it includes the MCU load rate, NPU load rate, and CDI value of each of the four indoor units (3 dimensions per unit, 12 dimensions in total), as well as the compressor frequency, condensing pressure, outdoor temperature, and current electricity price of the outdoor unit (4 dimensions), for a total of 16 dimensions; the return air temperature deviation has been incorporated into the CDI comprehensive calculation. The state vector is a complete description of the current operating state of the system, covering the computing load and comfort status of each indoor unit, as well as the refrigerant system status and environmental economic factors of the outdoor unit, serving as the input basis for optimization of the global scheduling model.

[0050] Action vector 'a' can be designed as 12-dimensional, for example including the inference frequency setpoint, task execution location code, and EEV opening adjustment amount for each of the four indoor units (3-dimensional for each unit, totaling 12 dimensions). The compressor frequency adjustment amount, as a system-level action, is synchronously output by the agent under the constraints of compressor frequency upper and lower limits and total refrigerant flow conservation. The inference frequency setpoint, task execution location code, and EEV opening adjustment amount correspond to computing power scheduling actions and cooling control actions, respectively, enabling the global scheduling model to synchronously adjust the computing power allocation and refrigerant flow of each indoor unit. The global scheduling model uses the comprehensive energy efficiency index as the reward objective, synchronously outputting the aforementioned computing power scheduling actions and cooling control actions through the Actor network. The Critic network evaluates the joint value, and the comprehensive energy efficiency and state parameter feedback form a new state, thereby achieving unified optimization of computing power collaboration and cooling control, reaching the optimal comprehensive energy efficiency.

[0051] The state vector is designed to be 16-dimensional, including 12 dimensions: the MCU load rate, NPU load rate, and CDI value of each of the four indoor units; and 4 dimensions: the outdoor unit's compressor frequency, condensing pressure, outdoor temperature, and current electricity price. Specifically, the MCU and NPU load rates are sampled by the local computing manager at 10-second intervals and normalized to the 0-1 range; the CDI is calculated locally and sampled at 10-second intervals, directly using its values ​​from -1 to +1; the compressor frequency, condensing pressure, and outdoor temperature are sampled by the outdoor unit's inverter, pressure sensor, and temperature sensor, respectively, and normalized to the 0-1 range using a min-max method; the current electricity price is derived from the electricity price signal and normalized to the 0-1 range. The state vector is sampled at 10-second intervals, with the global scheduling agent performing inference every 5 minutes, and additional sampling triggered by events such as changes in personnel status.

[0052] The action vector is designed in 12 dimensions, including the inference frequency setpoint, task execution location code, and EEV opening adjustment for each of the four indoor units. The inference frequency setpoint is a discrete action, with five levels: 0, 0.2, 0.5, 1, and 2 Hz. The task execution location code is also a discrete action, selecting one of three options: local, outdoor unit, or neighboring unit. The EEV opening adjustment is a continuous action, ranging from 0 to 100%. The compressor frequency adjustment is a system-level action, synchronously output by the agent while satisfying the constraints of compressor frequency upper and lower limits and total refrigerant flow conservation. The DDPG network outputs continuous actions, and the discrete levels are obtained through continuous value-to-level quantization mapping.

[0053] The reward function consists of a base term and a penalty term. The base term includes: a positive reward for energy efficiency improvement, a first negative reward for the proportion of computing power added to power consumption, and a second negative reward for exceeding comfort limits. Each term is multiplied by its corresponding weighting coefficient and then summed. The penalty term includes: a penalty based on the square of the superheat deviation when the evaporator outlet superheat is below 5°C, with a larger weight used to protect the compressor and system safety; a penalty based on the absolute value of the compressor frequency change rate to prevent surge and mechanical wear; and a penalty when the room temperature deviates from the set value and exceeds the bandwidth constraint.

[0054] To address the characteristics of air conditioning systems, such as high thermal inertia and delayed comfort feedback, the system is pre-trained in a digital twin environment, where thermodynamic time delays are explicitly modeled within the twin. Rewards are discounted and multi-step rewards are used to compensate for the delay. The current step uses readily measurable power consumption, frequency, and pressure as intermediate rewards, while comfort is injected through CDI moving average delay. The state vector extends the historical time window to help the agent perceive the time delay.

[0055] In digital twin simulations, the outdoor unit performs agent inference every 5 minutes, simultaneously outputting computing power scheduling and compressor adjustment actions. For example, in low-load nighttime scenarios, the agent decides to shut down the NPU power supply to the conference room and office areas A / B, retaining only local lightweight inference in the manager's office, reducing the additional power consumption to 0.3%. In extreme high-temperature afternoon scenarios, the agent decides to activate all NPUs, increasing the inference frequency to once every 2 seconds, and locking the compressor in the high-efficiency zone through refined refrigerant redistribution, resulting in a 12% improvement in system COP compared to traditional simulations. The outdoor unit records simulation operation logs and uses idle computing power for model fine-tuning during low-load nighttime periods. Incremental learning is triggered by the following conditions: the absolute value of CDI in any room exceeds 0.3 for more than 30 minutes; the predicted MAPE of heat load exceeds 15% within a 24-hour sliding window; and online concept drift detection determines a significant shift in feature distribution.

[0056] In some embodiments, after outputting the computing power scheduling instruction for the indoor unit to be optimized, the method further includes the following steps: Step 11: Maintain a credit score for each indoor unit, the credit score being used to assess the reliability of the indoor unit as an offloading target; the offloading target is the indoor unit that receives and executes the inference task of offloading other indoor units. Step 12: Sort the candidate unloading target sequence in descending order according to the effective computing power of each indoor unit, and send the candidate unloading target sequence to the indoor units; the effective computing power is calculated based on the indoor unit load rate and the credit score.

[0057] In this embodiment of the invention, the outdoor unit maintains a credit score for each indoor unit. This credit score is used to assess the reliability of the indoor unit as an offloading target. An offloading target refers to an indoor unit that receives and executes inference tasks offloaded by other indoor units. When an indoor unit lacks sufficient local computing power or the task requires merging data from adjacent rooms, it offloads the inference task to another indoor unit. The indoor unit receiving and executing this task is the offloading target. The credit score reflects the reliability of the indoor unit as an offloading target historically. The higher the score, the more reliable the indoor unit has been in previous offloading tasks, and the more suitable it is as a subsequent offloading target. The credit score is updated by the outdoor unit based on the offloading task execution results and heartbeat status when the indoor unit is an offloading target. For example, if the task is successfully completed and returned promptly, the score increases; if the return result verification fails or the heartbeat times out, the score decreases.

[0058] The outdoor unit generates a candidate unloading target sequence by sorting the indoor units in descending order of their effective computing power, and then distributes this sequence to each indoor unit. Effective computing power is a comprehensive indicator calculated based on the indoor unit's load rate and credit score, simultaneously measuring the indoor unit's idle capacity and historical reliability. A lower load rate indicates more idle computing power, and a higher credit score indicates better historical reliability. Sorting by effective computing power in descending order means that the indoor unit with the highest effective computing power is placed at the top of the sequence as the primary unloading target; the unit with the second highest effective computing power is placed second as a backup unloading target, and so on. After the outdoor unit distributes the candidate unloading target sequence to each indoor unit, when an indoor unit needs to unload an inference task, it can send an unloading request to the corresponding neighboring unit in sequence. If the current neighboring unit is busy or unresponsive, the request is forwarded to the next neighboring unit in the sequence; if all neighboring units in the sequence are unavailable, the process falls back to local degraded inference. The outdoor unit provides a unified reliability assessment and priority ranking basis for task offloading between indoor units, enabling indoor units to quickly and reliably select offloading targets when tasks need to be offloaded, avoiding inference failures or repeated offloading due to the selection of unreliable nodes, and improving the efficiency and reliability of computing power collaboration.

[0059] In some embodiments, the method further includes the following steps: Step 13: Update the credit score based on the execution result of the unloading task when the indoor unit is the unloading target and the heartbeat status of the indoor unit; Step 14: When the indoor unit's credit score is lower than a preset threshold, stop using the indoor unit as the unloading target, issue a stability alarm for the indoor unit, and trigger the indoor unit to perform a self-test; after the indoor unit passes the self-test, restore the indoor unit's credit score to the preset value.

[0060] In this embodiment of the invention, after maintaining its credit score, the outdoor unit continuously updates its credit score based on the execution results of the unloading task and the heartbeat status of the indoor unit, and takes corresponding measures when the credit score is too low. The outdoor unit updates the credit score of the indoor unit based on the execution results of the unloading task when the indoor unit is the unloading target and the heartbeat status of the indoor unit. The execution results of the unloading task reflect the quality of task completion when the indoor unit is the unloading target, including whether the task was successfully completed, whether the return time was timely, and whether the return result verification passed. The heartbeat status of the indoor unit reflects whether the indoor unit is online and responding normally, including whether it has failed to send computing power heartbeat frames for a preset number of consecutive times. When the unloading task is successfully completed and the return time is less than the expected time limit, it indicates that the indoor unit is reliable, and the credit score increases. When there is a task execution abnormality, a heartbeat abnormality, or the indoor unit actively rejects the unloading request, it indicates that the reliability of the indoor unit decreases, and the credit score is deducted. Among these, task execution abnormalities include failure to verify the return result; heartbeat abnormalities include failing to send computing power heartbeat frames to the outdoor unit for a preset number of consecutive times; and actively rejecting the unloading request includes returning a rejection indication to the requesting indoor unit because the local task has a higher priority. When points are deducted, the outdoor unit simultaneously records the reason for the deduction and a timestamp, which facilitates subsequent tracing and analysis.

[0061] When the credit score of an indoor unit falls below a preset threshold, the outdoor unit suspends its use as an offloading target, meaning it is no longer included in the candidate offloading target sequence to avoid offloading inference tasks to unreliable nodes. Simultaneously, the outdoor unit sends a stability alarm to the indoor unit and triggers a self-test. The self-test includes NPU self-test vectors, Flash verification, and sensor validity checks to identify any hardware or software malfunctions. After the indoor unit passes the self-test, the outdoor unit restores its credit score to the preset value, making it eligible again as an offloading target and placing it in an observation period. During this observation period, the credit score gradually increases after a preset number of successful task executions. The outdoor unit dynamically assesses the reliability of each indoor unit as an offloading target, promptly removing unreliable nodes and triggering their self-test recovery, avoiding inference failures or repeated offloading due to unreliable node selection, thus improving the reliability and stability of computing power collaboration.

[0062] In some embodiments, the unloading task execution result includes whether the inference task of unloading other indoor units by the indoor unit was successfully completed, the return time, and whether the return result verification passed; step 13 may include the following sub-steps: When an indoor unit successfully completes the unloading task of other indoor units and the return time is less than the expected time limit, the credit score is increased; when a task execution abnormality, heartbeat abnormality, or indoor unit actively refuses the unloading request occurs, the credit score is deducted; the task execution abnormality includes failure to verify the return result; the heartbeat abnormality includes the indoor unit failing to send a computing power heartbeat frame to the outdoor unit for a preset number of consecutive times.

[0063] In this embodiment of the invention, the unloading task execution result includes whether the indoor unit successfully completed the inference task of unloading other indoor units, the return time, and whether the return result verification passed. When an indoor unit successfully completes the unloading task of other indoor units and the return time is less than the expected time limit, it indicates that the indoor unit is reliable as an unloading target and responds promptly, and the outdoor unit increases the credit score of the indoor unit. When there is a task execution abnormality, heartbeat abnormality, or the indoor unit actively rejects the unloading request, the outdoor unit deducts the credit score of the indoor unit. Among them, the task execution abnormality includes the failure of the return result verification, such as the return result showing NaN or exceeding the physical range, indicating that the inference result of the indoor unit is unreliable; the heartbeat abnormality includes the indoor unit failing to send computing power heartbeat frames to the outdoor unit for a preset number of consecutive times, indicating that the indoor unit may be offline or have communication abnormalities; actively rejecting the unloading request includes the indoor unit returning a rejection indication to the requesting indoor unit because its own task has a higher priority, indicating that the indoor unit currently does not have the conditions to receive the unloading task. When the score is deducted, the outdoor unit simultaneously records the reason for the deduction and the timestamp for subsequent traceability and analysis. The outdoor unit can dynamically and precisely update the credit score of each indoor unit based on the actual execution of the unloading task and the heartbeat status of the indoor unit, so that the credit score truly reflects the reliability of each indoor unit as an unloading target and provides an accurate basis for selecting unloading targets.

[0064] For example, the outdoor unit maintains a credit score (Credit_i) for each indoor unit, with an initial value of 100 and a range of 0 to 100. The outdoor unit updates the credit score according to the rules based on the unloading task execution result and node heartbeat status: When the unloading task is successfully completed and the return time is less than the expected time limit, the score increases by 1, up to a maximum of 100, with no excessive rewards; when the return result verification fails (outputting NaN or exceeding the physical range), the score decreases by 5, triggering the result verifier's judgment; when three consecutive computing power heartbeats time out, the score decreases by 10, connecting with the node offline judgment mechanism; when actively rejecting an unloading request (because the local task has higher priority), the score decreases by 2, without excessive penalty, retaining flexibility. The score increment is denoted as Δ_i. When a score is deducted, the outdoor unit simultaneously records the reason for the deduction and a timestamp. The credit score is updated in the manner of Credit_i←clamp(Credit_i+Δ_i, 0, 100), meaning the updated score is limited to the range of 0 to 100.

[0065] When selecting unloading targets, the outdoor unit sorts them in descending order of effective computing power. Effective computing power Effective_i = (1 - Load_i) × (Credit_i / 100), where Load_i is the node's NPU load rate and Credit_i is the node's credit score. Effective computing power comprehensively considers the node's idle time and historical reliability, prioritizing the node with the highest effective computing power as the unloading target, and generating candidate chains accordingly.

[0066] When a node's credit score falls below 50, the outdoor unit suspends its use as an unloading target, issues a node stability alarm, and triggers a self-test for that node. The self-test includes NPU self-test vector, Flash verification, and sensor validity verification. After the self-test passes, the credit score recovers to 60 and enters an observation period. During the observation period, the score gradually increases after five consecutive successful task executions.

[0067] To further improve the accuracy of credit assessment, the credit score is decomposed into three sub-dimensions: reliability, timeliness, and data quality. Reliability R_i represents the historical uninstallation success rate; timeliness T_i represents the response delay compliance rate (the proportion of tasks with return times shorter than expected); and data quality Q_i represents the result verification pass rate. Example weights are w_r=0.5, w_t=0.3, and w_q=0.2. The three sub-dimensions are statistically analyzed using a sliding window, with the example window representing the most recent 50 tasks to ensure proximate causes are prioritized. The credit score can be calculated as Credit_i = 100 × (w_r·R_i + w_t·T_i + w_q·Q_i).

[0068] Credit scores are also linked to the inference arbitration mechanism: events judged as anomalous (result verification failed or re-verification failed) are penalized according to the above rules, while events with low confidence but ultimately verified as correct are not penalized, thus avoiding false positives. Furthermore, nodes with credit scores below 70 will have their latest model versions temporarily suspended, and will only receive stable versions after self-checking, to prevent anomalous nodes from contaminating the system model.

[0069] Reference Figure 2 This diagram illustrates a flowchart of another collaborative control method for a multi-split air conditioner provided by an embodiment of the present invention. The method is applied to an indoor unit, which is connected to an outdoor unit and other indoor units. The method may include the following steps: Step 201: Extract features from the collected environmental data to obtain a time-series feature vector, and upload the time-series feature vector to the outdoor unit; In this embodiment of the invention, the indoor unit is equipped with a sensor array, including a return air temperature sensor, a humidity sensor, a millimeter-wave human presence sensor, a CO2 sensor, and an ambient light sensor. Each sensor communicates with an I-type sensor. 2The indoor unit's MCU is connected via a C / SPI bus. The indoor unit collects environmental data through a sensor array, including temperature, humidity, human presence, CO2 concentration, and light intensity. It may also include time-coded and historical load data.

[0070] The indoor unit is equipped with a feature extraction network for an air conditioning heat load prediction model. This feature extraction network, consisting of an input layer to a hidden layer, extracts time-series feature vectors from raw sensor data. The indoor unit inputs the collected environmental data into this feature extraction network, which then extracts features from the multimodal environmental data and outputs a time-series feature vector. This time-series feature vector is an abstract feature representation of the raw environmental data, carrying information about the room's thermal environment.

[0071] After obtaining the time-series feature vector, the indoor unit uploads it to the outdoor unit via the indoor-outdoor unit collaborative communication bus. The outdoor unit deploys a load prediction network based on the same air conditioning heat load prediction model, which includes a hidden layer to an output layer. After receiving the time-series feature vectors uploaded by each indoor unit, the outdoor unit aggregates these vectors and inputs them into the load prediction network, which then outputs a global heat load prediction value. Through cross-node collaboration between feature extraction from the indoor unit's feature extraction network and global prediction from the outdoor unit's load prediction network, the inference process of the air conditioning heat load prediction model is distributed and completed collaboratively by both the indoor and outdoor units.

[0072] In some embodiments, the outdoor unit and the indoor unit are respectively deployed with pre-trained air conditioning heat load prediction models, the air conditioning heat load prediction models including feature extraction networks and load prediction networks; the global heat load prediction value is obtained by the outdoor unit based on the time-series feature vector of the load prediction network of the air conditioning heat load prediction model deployed on itself; step 201 may include the following sub-steps: Sub-step S41 involves inputting the collected environmental data into the feature extraction network and outputting the temporal feature vector through the feature extraction network.

[0073] In this embodiment of the invention, both the outdoor and indoor units are equipped with pre-trained air conditioning heat load prediction models, which include a feature extraction network and a load prediction network. The indoor unit uses the feature extraction network to extract time-series feature vectors from environmental data; the outdoor unit uses the load prediction network to output a global heat load prediction value based on the time-series feature vectors of each indoor unit. The global heat load prediction value is calculated by the outdoor unit based on its own deployed load prediction network, aggregating the time-series feature vectors uploaded by each indoor unit.

[0074] Under this premise, the indoor unit inputs the environmental data collected by the sensor array into its own deployed feature extraction network. This network extracts features from multimodal environmental data such as temperature, humidity, human presence, CO2 concentration, and light intensity, outputting a time-series feature vector. This time-series feature vector is an abstract feature representation of the original environmental data, carrying the state information of the room's thermal environment. After obtaining the time-series feature vector, the indoor unit uploads it to the outdoor unit via the indoor-outdoor unit collaborative communication bus, allowing the outdoor unit's load prediction network to perform global heat load prediction. Through the cross-node division of labor between the feature extraction network and the load prediction network, the inference process of the air conditioning heat load prediction model is distributed among the indoor and outdoor units for collaborative completion. By performing feature extraction locally and only uploading the time-series feature vector to the outdoor unit, the model inference process is distributed among the indoor and outdoor units, eliminating reliance on the cloud or a separate group controller, reducing communication latency, and improving system real-time performance.

[0075] Step 202: Calculate the comfort deviation based on the environmental data and the time-series feature vector; In this embodiment of the invention, the indoor unit extracts four components from environmental data: temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value. These components are then combined with the time-series feature vector output from the first layer of slicing to weightedly synthesize the comfort deviation. Specifically, the temperature deviation reflects the difference between the actual room temperature and the set temperature; the humidity deviation reflects the difference between the actual room humidity and the comfortable humidity range; the carbon dioxide concentration deviation reflects the difference between the room CO2 concentration and the comfortable concentration threshold; and the human body heat dissipation contribution value is determined based on the presence and activity intensity of people in the environmental data, characterizing the contribution of heat emitted by people in the room to the room's heat load. The greater the activity level of people, the greater the heat dissipation contribution, and the more likely the room is to be overheated.

[0076] The indoor unit weights and synthesizes the four components according to preset weighting coefficients to obtain the basic comfort deviation. Then, a trend correction is applied to the basic comfort deviation based on the time-series feature vector to obtain the final comfort deviation. The time-series feature vector contains information such as historical load and time coding, which can be used to determine the changing trend of the comfort deviation. For example, if it is predicted that the load will increase in the future, the comfort deviation can be fine-tuned in advance. The comfort deviation is a scalar value; a negative value indicates that it is too cold, and a positive value indicates that it is too hot. The larger the absolute value, the further it deviates from the comfort zone.

[0077] In some embodiments, step 202 may include the following sub-steps: Sub-step S51: Determine the temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value based on the environmental data; the human body heat dissipation contribution value characterizes the contribution of the heat emitted by a person in the room to the room's heat load. Sub-step S52: Determine the weighting coefficients corresponding to the temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value based on the time-series feature vector; Sub-step S53: The comfort deviation is obtained by weighting and synthesizing the temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value and corresponding weighting coefficients.

[0078] This implementation maps the three states output by the millimeter-wave human presence sensor—static, slightly moving, and significantly moving—to activity intensity levels and thermal voting contributions. Static corresponds to activity intensity level 0, with a metabolic equivalent of approximately 1.0 and a thermal voting contribution of 0. Slight movement (such as office work or typing) corresponds to activity intensity level 1, with a metabolic equivalent of approximately 1.2 to 1.5 and a thermal voting contribution of +0.2 to +0.4. Significant movement (such as walking or exercising) corresponds to activity intensity level 2, with a metabolic equivalent of approximately 2.0 to 3.0 and a thermal voting contribution of +0.5 to +1.0. The human body heat dissipation contribution value is estimated using the formula TV=tanh(0.3·L_act+0.7·R_occ)·I_present, where L_act is the activity intensity level, R_occ is the percentage of time the person is present within the current sliding window, and I_present is the presence indicator (1 for people in the room, 0 for people out of the room). This estimated value is weighted by temperature deviation, humidity deviation, and CO2 concentration deviation to synthesize the CDI, with an output range of -1 to +1.

[0079] The constraints for global optimization of the outdoor unit are as follows: The outdoor unit collects local decision broadcast frames from all indoor units every 30 seconds, and performs global consistency verification and refrigerant system constraint optimization. The constraints include four aspects: First, the compressor frequency upper and lower limits are 30 to 120 Hz; second, total refrigerant flow is conserved, meaning the sum of the opening degrees of each EEV is limited by the compressor displacement; third, superheat protection is implemented, meaning the evaporator outlet superheat is not lower than 5°C; and fourth, computing power load balancing is achieved, meaning the difference in NPU load rate among nodes is less than 30%. The outdoor unit operates based on a constraint reinforcement learning-based global scheduling model, outputting the target opening degree of each room's EEV, the target compressor frequency, and computing power scheduling instructions for each node. Through the above constraints and the global scheduling model, the outdoor unit achieves global optimization with the goal of optimal overall energy efficiency, while ensuring refrigerant system safety, meeting comfort standards, and achieving computing power load balancing.

[0080] In this embodiment of the invention, the indoor unit determines four components based on collected environmental data: temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution. Temperature deviation reflects the difference between the actual room temperature and the set temperature; humidity deviation reflects the difference between the actual room humidity and the comfortable humidity range; carbon dioxide concentration deviation reflects the difference between the room CO2 concentration and the comfortable concentration threshold; and the human body heat dissipation contribution represents the contribution of heat dissipated by people in the room to the room's heat load. This contribution is determined based on the presence and activity intensity of people in the environmental data; the greater the activity level, the greater the heat dissipation contribution, and the more likely the room is to be overheated. These four components reflect the room's thermal comfort state from four dimensions: temperature, humidity, air quality, and human body heat dissipation.

[0081] The indoor unit determines the weighting coefficients for temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value based on the time-series feature vector output from the first-layer segment. The time-series feature vector contains information such as historical load and time encoding, reflecting the degree of influence of each component on comfort under current operating conditions. For example, when human activity increases, the weight of the human body heat dissipation contribution value can be increased accordingly; when the temperature change trend is obvious, the weight of the temperature deviation can be increased accordingly. By dynamically determining the weighting coefficients of each component based on the time-series feature vector, the calculation of comfort deviation can adapt to changes in the dominant comfort factors under different operating conditions.

[0082] The indoor unit calculates the comfort deviation by weighting and synthesizing temperature deviation, humidity deviation, carbon dioxide concentration deviation, and the contribution of human body heat dissipation, along with their corresponding weighting coefficients. The comfort deviation is a scalar value; a negative value indicates it's too cold, a positive value indicates it's too hot, and a larger absolute value indicates it deviates further from the comfort zone. This comfort deviation, along with the current fan speed, is then encapsulated into a local decision broadcast frame and broadcast on the indoor-outdoor unit collaborative communication bus for use by the outdoor unit in global optimization. The indoor unit extracts four dimensions of comfort components from environmental data and dynamically adjusts the weights of each component based on temporal feature vectors, then weights and synthesizes them to obtain the comfort deviation, achieving a multi-dimensional, adaptive assessment of room comfort.

[0083] Step 203: Obtain the current fan speed, and encapsulate the comfort deviation and the current fan speed into a local decision broadcast frame for broadcasting; In this embodiment of the invention, after completing the comfort deviation calculation, the indoor unit reads the current fan speed from the fan drive module. This current fan speed is the actual operating speed of the indoor unit's fan, reflecting the current air supply status of the indoor unit. The indoor unit encapsulates the calculated comfort deviation and the current fan speed together into a local decision broadcast frame, which is then broadcast on the indoor-outdoor unit collaborative communication bus at a broadcast period.

[0084] Local decision broadcast frames serve as the carriers for indoor units to feed back local decision information to outdoor units. By broadcasting local decision broadcast frames, indoor units inform outdoor units of the comfort deviation of their respective rooms and the current fan speed of their units, enabling outdoor units to obtain the comfort status of each room and the operating status of each indoor unit. Outdoor units collect all local decision broadcast frames from indoor units during the global optimization cycle, extracting the comfort deviation and current fan speed of each indoor unit as local state inputs for global optimization.

[0085] Compared to existing technologies where indoor units only report limited parameters such as return air temperature, this solution's local decision broadcast frames from indoor units include richer local decision information such as comfort deviation and current fan speed. Comfort deviation reflects the current thermal comfort state of each room and serves as the basis for global optimization to determine which rooms need adjustment and in what direction. Current fan speed reflects the current operating status of each indoor unit and is a reference for global optimization calculations of refrigerant flow redistribution and fan speed adjustment. By collecting these local decision broadcast frames, the outdoor unit can balance the comfort needs of each room with the operating status of each indoor unit in global optimization, achieving more refined zone control.

[0086] Step 204: Receive the recommended control command sent by the outdoor unit; In this embodiment of the invention, after completing global optimization, the outdoor unit generates a recommended control command corresponding to the indoor unit to be optimized in this round, and sends the recommended control command to the corresponding indoor unit to be optimized. The indoor unit receives the recommended control command sent by the outdoor unit through the indoor-outdoor unit collaborative communication bus.

[0087] The recommended control command is generated by the outdoor unit after performing global optimization with the goal of achieving optimal overall energy efficiency while meeting refrigerant constraints. It includes the target opening degree of the electronic expansion valve, the target frequency of the compressor, and a computing power scheduling command. Specifically, the target opening degree of the electronic expansion valve instructs the indoor unit to adjust its electronic expansion valve to the target opening degree to regulate the refrigerant flow to the corresponding indoor unit; the target frequency of the compressor, as a system-level command, is used to adjust the total refrigerant circulation volume of the system; and the computing power scheduling command instructs the indoor unit to adjust the inference task execution position or inference frequency. For indoor units to be optimized, the outdoor unit issues the recommended control command; for indoor units not to be optimized, the outdoor unit does not issue new recommended control commands, and the current control commands for non-optimized indoor units remain unchanged.

[0088] Step 205: Output local control commands based on the current indoor unit operating status, arbitrate the recommended control commands with the local control commands, generate target control commands based on the arbitration result, and execute the target control commands.

[0089] In this embodiment of the invention, the indoor unit MCU is pre-loaded with a lightweight rule inference engine, which is a fuzzy PID or decision tree. After receiving the recommended control command issued by the outdoor unit, the indoor unit outputs a local control command based on the current operating status of the indoor unit by the local rule engine. The local control command is a conservative control value given by the rule engine based on the current operating status, which does not rely on inference and has high reliability.

[0090] The indoor unit arbitrates the recommended control commands issued by the outdoor unit with the local control commands output by the local rule engine using heterogeneous redundant inference based on model confidence. The arbitration comprehensively considers the confidence of each of the two control commands and the deviation between them: if the confidence of both commands is higher than the first confidence threshold and the deviation is less than the first deviation threshold, then the corresponding control quantities in the two control commands are weighted and averaged to obtain the target control command, using the confidence of each command as a weight; if either command is lower than the second confidence threshold, then the local control command output by the local rule engine is used as the target control command, and a low-confidence event is recorded; if the deviation between the two commands is greater than the second deviation threshold, then regardless of the confidence level, the local control command is used as the target control command, and a model verification task is triggered.

[0091] After generating the target control command based on the arbitration result, the indoor unit executes the target control command, that is, it controls the corresponding actuators to act according to the target opening degree of the electronic expansion valve and the target speed of the fan in the target control command. After receiving the recommended control command from the outdoor unit, the indoor unit does not execute it directly, but first arbitrates it with the conservative control command output by the local rule engine. If the inference result is reliable, the fused result is used; if the inference result is abnormal, it is downgraded to the rule engine output. In this way, while ensuring control accuracy, it avoids erroneous commands from directly acting on the compressor and electronic expansion valve due to inference errors, thereby improving the safety and reliability of the system.

[0092] For example, the indoor unit arbitrates the recommended control commands issued by the outdoor unit with the local control commands output by the local rule engine. If Conf_npu > 0.85 and Conf_od > 0.85, and the deviation between the two results is less than 10%, a confidence-weighted average fusion value is used. The fusion weight is taken from the confidence levels of the two sources, with the inference result having a higher confidence level having a greater weight in the fusion value. The fusion formula is Out = (Conf_npu·Out_npu + Conf_od·Out_od) / (Conf_npu + Conf_od). If either confidence level is lower than 0.7, the local control commands output by the MCU rule engine are used, and a low-confidence event is recorded. If the deviation between the two results is greater than 20%, regardless of the confidence level, the local control commands output by the MCU rule engine are used, and a model verification task is triggered, with the outdoor unit issuing test vectors to re-verify the two NPUs. The arbitration result is issued and executed as the final EEV opening and fan speed commands.

[0093] To avoid frequent switching at threshold critical points, arbitration employs a state machine with hysteresis. Arbitration states include normal, low confidence, conflict, and degraded states, and state switching requires confirmation over two consecutive control cycles. In conflict or degraded states, in addition to using the rule engine output, the compressor frequency change rate (exemplarily not exceeding 2Hz / control cycle) and the single-step change in EEV opening degree are simultaneously limited to prevent drastic actions caused by abnormal commands. Since quantization errors exist after INT8 quantization compression of feature vectors, the outdoor unit sends confidence calibration parameters along with the scaling factor during model synchronization. This performs offline calibration of the quantized confidence threshold distribution to avoid frequent degrades due to systematic low confidence caused by quantization.

[0094] The outdoor unit distributes test vectors covering typical operating conditions and boundary samples to re-examine the inference outputs of the two NPUs. If the indoor unit's NPU is determined to be abnormal, the node reverts to the MCU rule engine output and alerts the outdoor unit, while simultaneously triggering a self-check on that node. If the outdoor unit's NPU is determined to be abnormal, each indoor unit switches to the local rule engine's backup mode, the outdoor unit marks the model as abnormal, and rolls back to the previous stable version. All arbitration events (time, result deviation, confidence level, adopted path) are uploaded to the outdoor unit log for subsequent model fine-tuning, anomaly diagnosis, and credit score updates, and are linked with the computing power node credit score scheduling mechanism. In the exemplary scenario, this mechanism can intercept all erroneous inference commands, significantly reducing the number of abnormal compressor frequency jumps and significantly reducing overheat protection malfunctions.

[0095] After the inference results generate control commands, the impact of chip power consumption and communication power consumption on system energy efficiency is comprehensively calculated. Each indoor unit independently calculates the Comfort Deviation Index (CDI) for its room, based on local sensor data and the output of the first-layer model, with a 10-second cycle. The CDI comprehensively considers temperature deviation, humidity deviation, human thermal voting (estimated through human presence duration and activity intensity), and CO2 concentration, and outputs a scalar value ranging from -1 to +1, with negative values ​​indicating coldness and positive values ​​indicating heatness. The indoor unit encapsulates the CDI value, current fan speed, and local computing power margin into a local decision broadcast frame, which is broadcast on the bus with a 50ms cycle. Through the above arbitration mechanism, after receiving the recommended control command from the outdoor unit, the indoor unit does not execute it directly, but rather conducts joint arbitration with the local control command output by the local rule engine based on confidence and deviation. When the inference result is reliable and consistent with the rule engine output, the fused result is adopted; when the inference result is abnormal or conflicts with the rule engine output, it is downgraded to the local control command, thereby improving the system's safety and reliability while ensuring control accuracy.

[0096] In some embodiments, step 205 may include the following sub-steps: Sub-step S61: Obtain the first confidence level corresponding to the recommended control instruction and the second confidence level corresponding to the local control instruction, and calculate the deviation between the corresponding control quantities in the recommended control instruction and the local control instruction; Sub-step S62: If the first confidence level is higher than the first confidence level threshold, the second confidence level is higher than the first confidence level threshold, and the deviation is less than the first deviation threshold, then the first confidence level and the second confidence level are used as weights to perform a weighted average of the corresponding control quantities in the recommended control command and the local control command to obtain the target control command. Sub-step S63: If the first confidence level or the second confidence level is lower than the second confidence level threshold, then the local control instruction is used as the target control instruction; the first confidence level threshold is greater than the second confidence level threshold. Sub-step S64: If the deviation is greater than the second deviation threshold, then the local control command is used as the target control command; the first deviation threshold is less than the second deviation threshold.

[0097] In this embodiment of the invention, the indoor unit acquires the first confidence level corresponding to the recommended control command issued by the outdoor unit and the second confidence level corresponding to the local control command output by the local rule engine, and calculates the deviation between the corresponding control quantities in the recommended control command and the local control command. The first confidence level reflects the reliability of the outdoor unit's inference result, the second confidence level reflects the reliability of the local rule engine's output, and the deviation reflects the magnitude of the difference between the two control commands on the same control quantity.

[0098] If the first confidence level is higher than the first confidence level threshold, the second confidence level is higher than the first confidence level threshold, and the deviation is less than the first deviation threshold, it indicates that both control commands are reliable and mutually corroborative. Then, using the first and second confidence levels as weights, a weighted average is calculated on the corresponding control quantities in the recommended control command and the local control command to obtain the target control command. The higher the confidence level, the greater the weight of the corresponding control command in the weighted average.

[0099] If either the first confidence level or the second confidence level is lower than the second confidence level threshold, it indicates that at least one control command is unreliable, and the local control command is then used as the target control command. The first confidence level threshold is greater than the second confidence level threshold, forming a hierarchical judgment between high-confidence and low-confidence regions.

[0100] If the deviation is greater than the second deviation threshold, it indicates that the two control commands differ too much and cannot be mutually verified. In this case, regardless of the confidence level, the local control command is used as the target control command. The first deviation threshold is less than the second deviation threshold, forming a hierarchical judgment between the small deviation zone and the large deviation zone.

[0101] The indoor unit performs joint arbitration between recommended control commands and local control commands based on confidence level and deviation. When the inference result is reliable and consistent with the rule engine output, the fused result is adopted. When the inference result is abnormal or conflicts with the rule engine output, it is downgraded to the local control command, thereby improving the system's safety and reliability while ensuring control accuracy.

[0102] In some embodiments, refer to Figure 3 The diagram illustrates a flowchart of a collaborative control method for multi-split air conditioners according to an embodiment of the present invention, specifically a power unloading decision process. The method may further include the following steps: like Figure 3 As shown in the figure, this illustrates the complete process of indoor unit computing power offloading decision-making in this embodiment. After the indoor unit sensors collect environmental data, they trigger an inference task. The local computing power manager first evaluates the current local load rate and the real-time requirements of the task to be executed, determining whether the local load rate is lower than a preset threshold and whether the task completion time is less than a preset time threshold. If the conditions are met, it indicates that the local computing power is sufficient and the task has high real-time requirements, so the inference task is executed directly locally; otherwise, the offloading process begins.

[0103] In the unloading process, the indoor unit first obtains the candidate unloading target sequence generated by the outdoor unit in descending order based on the comprehensive score. The comprehensive score takes into account the idle computing power index, task type matching degree, communication latency, power state factor, and historical unloading success rate. Then, it sends unloading request frames to the corresponding neighboring units in the candidate sequence. If the current neighboring unit experiences a load increase and returns a busy indication during inference, or fails to respond after a timeout, the system switches to the next neighboring unit along the candidate sequence. If all neighboring units in the candidate sequence are unavailable, the system falls back to local degradation inference, retaining the simplest model to complete the inference locally. Regardless of whether the inference is executed locally, successfully unloaded, or falls back to degradation inference, control commands are ultimately generated.

[0104] Step 21: Obtain the current local load rate and the real-time requirements of the inference task to be executed; Step 22: If the current local load rate is lower than a preset load rate threshold and the required completion time of the real-time requirement of the inference task to be executed is less than a preset time threshold, then the inference task to be executed is executed locally. Step 23: If the current local load rate is greater than or equal to the preset load rate threshold or the inference task to be executed needs to integrate data from other indoor units, then the inference task to be executed is sent to the outdoor unit or other indoor units with load rates lower than the preset load rate threshold through the offload protocol.

[0105] In this embodiment of the invention, the indoor unit acquires the current local load rate and the real-time requirements of the inference task to be executed. The current local load rate reflects the current computing power occupancy of the indoor unit's MCU and NPU, and is obtained by the local computing power manager; the real-time requirements of the inference task to be executed reflect how long the task needs to be completed, characterized by the required completion time. The shorter the required completion time, the higher the real-time requirement.

[0106] If the current local load rate is lower than the preset load rate threshold, it indicates that the local computing power is sufficient; and if the required completion time of the inference task to be executed is less than the preset time threshold, it indicates that the task has high real-time requirements and is not suitable for round-trip communication; then the indoor unit executes the inference task to be executed locally. Local execution does not require transmission through the indoor-outdoor unit collaborative communication bus, has the lowest latency, and is suitable for tasks with high real-time requirements and sufficient local computing power.

[0107] If the current local load rate is greater than or equal to the preset load rate threshold, it indicates insufficient local computing power; or if the inference task to be executed requires the fusion of data from other indoor units, it indicates that the local data is incomplete and cannot be completed independently. In this case, the indoor unit sends the inference task to the outdoor unit or another indoor unit with a load rate lower than the preset load rate threshold through the offloading protocol. Offloading to the outdoor unit is suitable for scenarios where the task requires global data or needs to be processed by the outdoor unit's NPU; offloading to other indoor units is suitable for scenarios where neighboring units have surplus computing power. The indoor unit adaptively selects between local execution and task offloading based on the local load rate and the task's real-time requirements. This ensures the response speed of high-real-time tasks while also allowing tasks to be handed over to other nodes for execution when local computing power is insufficient or data is incomplete, thus achieving dynamic allocation of computing resources.

[0108] In some embodiments, step 23 may include the following sub-steps: Obtain the candidate unloading target sequence sent by the outdoor unit; the candidate unloading target sequence is generated by the outdoor unit in descending order of effective computing power, and the effective computing power is calculated based on the indoor unit load rate and credit score; the inference task to be executed is taken as an unloading request frame and sent to the corresponding indoor unit in sequence according to the candidate unloading target sequence; the unloading request frame includes task type, input data digest, expected completion time limit and priority level; if the current indoor unit increases its load and returns a busy indication during inference, the next candidate indoor unit is determined along the candidate unloading target sequence; if all candidate indoor units in the candidate unloading target sequence return a busy indication or do not respond after timeout, the process falls back to local degraded inference.

[0109] The computing power topology table is maintained by the outdoor unit. After the system powers on, the outdoor unit broadcasts a computing power discovery frame, and each indoor unit responds and reports its own hardware computing power specifications. The outdoor unit summarizes the hardware computing power specifications reported by each indoor unit and generates the system computing power topology table. This table records the NPU load rate, available memory, supported model version, power status, and network topology location of each node, and is updated with the computing power heartbeat frames periodically reported by each indoor unit.

[0110] The computing power topology table maintained by the outdoor unit provides the data foundation for selecting unloading targets for the indoor unit. When an indoor unit needs to perform task unloading, the local computing power manager selects the unloading target based on the comprehensive score. The comprehensive score is: Score(i) = w1·C_idle(i) + w2·M_match(i) + w3·(1 / D(i)) + w4·E_power(i) + w5·H_succ(i). Where, C_idle(i) is the idle computing power index, which is equal to 1 minus the NPU load rate; M_match(i) is the task type matching degree, indicating whether the target node holds the model shards and operators required for the task, with a value of 0 or 1; D(i) is the communication latency between the requesting node and the target node; E_power(i) is the power state factor, which is 1 for grid-powered nodes and 0.5 for battery-powered or low-power nodes; H_succ(i) is the historical unloading success rate. The weights w1 to w5 are 0.4, 0.2, 0.2, 0.1, and 0.1, respectively. When candidate nodes have the same score, they are selected deterministically by sorting them by node ID and a candidate chain is generated. That is, if the first choice fails, the second-best node is selected in turn.

[0111] The indoor units encapsulate the inference tasks to be executed into offload request frames according to the candidate offload target sequence and send them to the corresponding indoor units. The offload request frame carries the task type, input data digest, expected completion time limit, priority level, lifespan, and expected completion time limit. If the target node's load increases during inference, for example, if the heartbeat reported NPU load rate exceeds 80%, the target node returns a busy indication to the requesting node, and the requesting node switches to the next node along the candidate chain; if all nodes in the candidate chain return busy indications or do not respond after timeout, the process falls back to local degraded inference, that is, the simplest model is retained to complete the inference locally, realizing task migration and rollback, and ensuring end-to-end real-time performance.

[0112] When an indoor unit needs to unload an inference task, it selects an unloading target based on the system computing power topology table and comprehensive score maintained by the outdoor unit. It tries the candidate chain in sequence, and when the target node is busy, it switches to the next node. When all nodes are unavailable, it rolls back to local degradation inference. This avoids inference failure due to the selection of unreliable nodes and ensures that the task can still be completed in the worst case. It realizes dynamic allocation of computing power resources and end-to-end real-time guarantee.

[0113] In this embodiment of the invention, the indoor unit acquires a sequence of candidate unloading targets sent by the outdoor unit. This sequence is generated by the outdoor unit, sorted in descending order of effective computing power. Effective computing power is calculated based on the indoor unit's load rate and credit score. Higher effective computing power indicates a greater degree of idle time and better historical reliability of the indoor unit, making it more suitable as an unloading target. Following the order of the candidate unloading target sequence, the indoor unit sequentially encapsulates the inference task to be executed into an unloading request frame and sends it to the corresponding indoor unit. The unloading request frame includes the task type, input data digest, expected completion time limit, and priority level, enabling the receiver to understand the nature of the task, data source, time limit requirements, and priority, and thus determine whether to accept the unloading request.

[0114] If the current indoor unit experiences a load increase and returns a busy indication during inference, it means that the indoor unit is currently unable to receive offload tasks. The requesting indoor unit then determines the next candidate indoor unit in the candidate offload target sequence and sends an offload request to it. If all candidate indoor units in the candidate offload target sequence return a busy indication or time out without response, it means that there are currently no available offload targets. The requesting indoor unit then falls back to local degraded inference, i.e., it uses the simplest model to complete the inference locally, ensuring uninterrupted inference output.

[0115] When an indoor unit needs to unload an inference task, it attempts to unload the task sequentially according to the candidate unload target sequence maintained by the outdoor unit to avoid inference failure due to the selection of an unreliable node. If all neighboring units in the candidate sequence are unavailable, it falls back to local degraded inference to ensure that the task can still be completed in the worst case.

[0116] In some embodiments, the method further includes: If the loss of the master node heartbeat of the outdoor unit is detected, it is determined that the communication between the indoor unit and the outdoor unit is interrupted, and the local control command is executed.

[0117] In this embodiment of the invention, the indoor unit also performs communication interruption detection and local autonomy. If the indoor unit detects that the master node heartbeat of the outdoor unit has been lost, it determines that the communication between the indoor unit and the outdoor unit has been interrupted. At this time, the indoor unit no longer waits for the recommended control commands issued by the outdoor unit, but directly executes the local control commands, that is, the conservative control commands output by the local rule engine according to the current operating status of the indoor unit, in order to maintain the basic operation of the indoor unit.

[0118] The indoor unit periodically listens for master node heartbeat frames broadcast by the outdoor unit on the indoor-outdoor unit collaborative communication bus. If no master node heartbeat frames are received for a preset number of consecutive cycles, it is determined that the master node heartbeat has been lost and communication has been interrupted. In this case, the indoor unit enters emergency autonomous mode, independently adjusting the fan speed according to the local comfort deviation, and maintaining the electronic expansion valve opening at the last value before the fault, ensuring that basic cooling or heating is not interrupted. The local control commands output by the local rule engine are directly executed as target control commands, without needing to arbitrate with the recommended control commands issued by the outdoor unit, because the outdoor unit is unreachable at this time and the recommended control commands cannot be received. After detecting a communication interruption with the outdoor unit, the indoor unit can automatically switch to local autonomous mode, relying on the local rule engine and local comfort deviation to maintain basic control, avoiding indoor unit shutdown due to outdoor unit failure or communication interruption, thus improving the system's reliability and robustness.

[0119] Reference Figure 4 The diagram shows a flowchart of a collaborative control method for a multi-split air conditioner according to an embodiment of the present invention. Figure 4 This illustrates the complete process of an embodiment of the present invention, such as... Figure 4 As shown, the indoor unit collects environmental data, extracts time-series feature vectors through its own deployed feature extraction network, and uploads them to the outdoor unit. Simultaneously, the indoor unit calculates the comfort deviation, obtains the current fan speed, and encapsulates both into a local decision broadcast frame, which is then broadcast on the indoor-outdoor unit collaborative communication bus. The outdoor unit receives the time-series feature vectors uploaded by each indoor unit, aggregates them, inputs them into its own deployed load prediction network, and outputs a global heat load prediction value. Subsequently, based on the global heat load prediction value, it generates recommended refrigerant flow levels and recommended fan speeds for each indoor unit, and collects all the local decision broadcast frames broadcast by the indoor units to obtain the comfort deviation, current fan speed, and local computing power margin for each indoor unit.

[0120] After collecting the above information, the outdoor unit identifies the indoor units requiring optimization in this round. Under refrigerant constraints, it performs global optimization with the goal of achieving optimal overall energy efficiency, generating recommended control commands and sending them to the corresponding indoor units. Upon receiving these commands, the indoor unit arbitrates with the local control commands output by the local rule engine based on the current operating status of the indoor unit. It comprehensively assesses the confidence and deviation of the two control commands, generates a target control command based on the arbitration result, and executes it. Through this process, the inference process of the air conditioning heat load prediction model is completed collaboratively by the indoor and outdoor units. Global optimization only adjusts the indoor units to be optimized, and the indoor units execute the sent recommended control commands after local arbitration.

[0121] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0122] Reference Figure 5 The diagram shows a structural block diagram of a collaborative control system for a multi-split air conditioner provided by an embodiment of the present invention. The system includes an outdoor unit 301 and multiple indoor units 302, and the indoor units 302 are respectively connected to the outdoor unit 301 and other indoor units 302. The indoor unit 302 is used to extract features from the collected environmental data to obtain a time-series feature vector, and upload the time-series feature vector to the outdoor unit; The outdoor unit 301 is used to receive the time-series feature vectors uploaded by each indoor unit and determine the global heat load prediction value based on the time-series feature vectors of each indoor unit. The indoor unit 302 is used to calculate the comfort deviation based on the environmental data and the time-series feature vector; obtain the current fan speed; and encapsulate the comfort deviation and the current fan speed into a local decision broadcast frame for broadcasting. The outdoor unit 301 is used to generate recommended refrigerant flow rate and recommended fan speed for each indoor unit based on the global heat load prediction value; collect local decision broadcast frames broadcast by all indoor units; and perform global optimization with the goal of optimal comprehensive energy efficiency under the condition of satisfying refrigerant constraints, based on the global heat load prediction value, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate and the recommended fan speed, to generate recommended control commands for each indoor unit and send the recommended control commands to the corresponding indoor unit. The indoor unit 302 is used to receive the recommended control command sent by the outdoor unit, output the local control command according to the current operating status of the indoor unit, arbitrate the recommended control command and the local control command, generate the target control command based on the arbitration result, and execute the target control command.

[0123] In some embodiments, the outdoor unit and the indoor unit are respectively deployed with pre-trained air conditioning heat load prediction models, the air conditioning heat load prediction models including feature extraction networks and load prediction networks; the time-series feature vectors are extracted by the indoor unit from the original sensor data based on the feature extraction network of the air conditioning heat load prediction model deployed by the indoor unit itself; The outdoor unit 301 is used to aggregate the time-series feature vectors of each indoor unit and input them into the load prediction network, and output the global heat load prediction value through the load prediction network.

[0124] In some embodiments, the outdoor unit 301 is configured to determine the total refrigerant demand and total air volume demand based on the global heat load forecast; obtain the rated capacity or historical load percentage of each indoor unit, and allocate recommended refrigerant flow to each indoor unit according to the rated capacity or historical load percentage of each indoor unit; the sum of the recommended refrigerant quantities of all indoor units is the total refrigerant demand; and determine the fan speed of each indoor unit based on the total air volume demand and the rated air volume of each indoor unit.

[0125] In some embodiments, the local decision broadcast frame further includes the local computing power margin of the corresponding indoor unit; The outdoor unit 301 is used to determine whether the comfort deviation of each indoor unit is within a preset deviation range, and to identify indoor units whose comfort deviation is not within the preset deviation range as indoor units to be optimized; and / or, to perform a computing power load balancing check based on the local computing power margin of each indoor unit, and to identify indoor units whose load rate deviates from the average load rate by more than a preset load rate deviation threshold as indoor units to be optimized; based on the global heat load prediction value, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate level and the recommended fan speed, to perform global optimization with the goal of optimal comprehensive energy efficiency under the condition of satisfying refrigerant constraints, to generate recommended control instructions corresponding to the indoor units to be optimized, and to send the recommended control instructions to the corresponding indoor units to be optimized.

[0126] In some embodiments, the outdoor unit is equipped with a global scheduling model; the refrigerant constraints include compressor frequency being within the upper and lower limits, total refrigerant flow being conserved, evaporator outlet superheat not being lower than a preset value, and computing power load balancing; The outdoor unit 301 is used to construct a state vector, which includes the load rate, comfort deviation, and current fan speed of the indoor unit to be optimized, as well as the compressor frequency, condensing pressure, outdoor temperature, and current electricity price of the outdoor unit. The state vector, the global heat load prediction value, the recommended refrigerant flow rate level, and the recommended fan speed are input into the global scheduling model. Under the condition of satisfying refrigerant constraints, with the goal of optimal comprehensive energy efficiency, the global scheduling model outputs parameter adjustment instructions and computing power scheduling instructions for the indoor unit to be optimized. The parameter adjustment instructions are used to instruct the opening degree of the electronic expansion valve of the indoor unit to be optimized to be set to the target opening degree and the compressor frequency to be set to the target frequency. The computing power scheduling instructions are used to migrate some inference tasks of the indoor units with load rates higher than the preset load rate threshold to the indoor units with load rates lower than the preset load rate threshold, or adjust the inference frequency of the corresponding indoor units.

[0127] In some embodiments, the outdoor unit 301 is used to simultaneously output computing power scheduling actions and cooling control actions through the global scheduling model with the comprehensive energy efficiency index as the reward target; the comprehensive energy efficiency index is the ratio of cooling capacity to comprehensive power consumption, and the comprehensive power consumption includes compressor power consumption, operating power consumption of all chips in the indoor and outdoor units, and communication bus power consumption; the computing power scheduling actions include executing inference tasks locally, offloading inference tasks to the outdoor unit or other indoor units, and adjusting the frequency of inference tasks executed by the indoor units; the cooling control actions include compressor frequency and electronic expansion valve opening.

[0128] In some embodiments, after outputting the computing power scheduling instruction for the indoor unit to be optimized, the outdoor unit 301 is used to maintain a credit score for each indoor unit. The credit score is used to evaluate the reliability of the indoor unit as an offloading target. The offloading target is an indoor unit that receives and executes inference tasks offloading from other indoor units. A candidate offloading target sequence is generated by sorting the effective computing power of each indoor unit in descending order, and the candidate offloading target sequence is sent to the indoor units. The effective computing power is calculated based on the indoor unit load rate and the credit score.

[0129] In some embodiments, the outdoor unit 301 is used to update the credit score based on the execution result of the unloading task when the indoor unit is the unloading target and the heartbeat status of the indoor unit; when the credit score of the indoor unit is lower than a preset threshold, the unloading of the indoor unit is suspended, the stability alarm of the indoor unit is issued, and the indoor unit is triggered to perform a self-test; after the indoor unit passes the self-test, the credit score of the indoor unit is restored to the preset value.

[0130] In some embodiments, the unloading task execution result includes whether the inference task of the indoor unit executing the unloading of other indoor units was successfully completed, the return time, and whether the return result verification passed. The outdoor unit 301 is used to increase the credit score when the indoor unit successfully completes the unloading task of other indoor units and the return time is less than the expected time limit; and to deduct the credit score when there is a task execution abnormality, heartbeat abnormality or indoor unit actively refuses the unloading request; the task execution abnormality includes failure to verify the return result; the heartbeat abnormality includes the indoor unit failing to send computing power heartbeat frames to the outdoor unit for a preset number of consecutive times.

[0131] In some embodiments, the outdoor unit and the indoor unit are respectively deployed with pre-trained air conditioning heat load prediction models, the air conditioning heat load prediction models including feature extraction networks and load prediction networks; the global heat load prediction value is obtained by the outdoor unit based on the load prediction network of the air conditioning heat load prediction model deployed by itself, based on the time-series feature vector; The indoor unit 302 is used to input the collected environmental data into the feature extraction network and output the time-series feature vector through the feature extraction network.

[0132] In some embodiments, the indoor unit 302 is used to determine temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value based on the environmental data; the human body heat dissipation contribution value characterizes the contribution of heat emitted by a person in the room to the room's heat load; determine the weighting coefficients corresponding to the temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value based on the time-series feature vector; and perform weighted synthesis based on the temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value and the corresponding weighting coefficients to obtain the comfort deviation.

[0133] In some embodiments, the indoor unit 302 is configured to acquire a first confidence level corresponding to the recommended control command and a second confidence level corresponding to the local control command, and calculate the deviation between the corresponding control quantities in the recommended control command and the local control command; if the first confidence level is higher than a first confidence threshold, the second confidence level is higher than the first confidence threshold, and the deviation is less than a first deviation threshold, then the first confidence level and the second confidence level are used as weights to perform a weighted average of the corresponding control quantities in the recommended control command and the local control command to obtain the target control command; if the first confidence level or the second confidence level is lower than the second confidence threshold, then the local control command is used as the target control command; the first confidence threshold is greater than the second confidence threshold; if the deviation is greater than the second deviation threshold, then the local control command is used as the target control command; the first deviation threshold is less than the second deviation threshold.

[0134] In some embodiments, the indoor unit 302 is used to obtain the current local load rate and the real-time requirements of the inference task to be executed; if the current local load rate is lower than a preset load rate threshold and the required completion time of the real-time requirements of the inference task to be executed is less than a preset time threshold, then the inference task to be executed is executed locally; if the current local load rate is greater than or equal to the preset load rate threshold or the inference task to be executed needs to integrate data from other indoor units, then the inference task to be executed is sent to the outdoor unit or other indoor units with load rates lower than the preset load rate threshold through an offload protocol.

[0135] In some embodiments, the indoor unit 302 is used to obtain a candidate unloading target sequence sent by the outdoor unit; the candidate unloading target sequence is generated by the outdoor unit in descending order of effective computing power, and the effective computing power is calculated based on the indoor unit load rate and credit score; the inference task to be executed is an unloading request frame, which is sent to the corresponding indoor unit in sequence according to the candidate unloading target sequence; the unloading request frame includes task type, input data digest, expected completion time limit and priority level; if the current indoor unit's load increases during inference and returns a busy indication, the next candidate indoor unit is determined along the candidate unloading target sequence; if all candidate indoor units in the candidate unloading target sequence return a busy indication or time out without response, the process falls back to local degraded inference.

[0136] In some embodiments, the indoor unit 302 is configured to determine that communication between the indoor unit and the outdoor unit is interrupted if the loss of the master node heartbeat of the outdoor unit is detected, and to execute the local control command.

[0137] As the system implementation is basically similar to the method implementation, it is described in a relatively simple way. For relevant details, please refer to the description of the method implementation.

[0138] This invention also provides a multi-split air conditioner, including: a processor, a memory, and a computer program stored in the memory and capable of running on the processor. When the computer program is executed by the processor, it implements the various processes of the cooperative control method embodiment of the multi-split air conditioner as described above, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0139] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the cooperative control method embodiment for multi-split air conditioners as described above, and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0140] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0141] Furthermore, it should be noted that the scope of the methods and apparatus in the embodiments of the present invention is not limited to performing functions in the order shown or discussed. It may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0142] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0143] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A collaborative control method for multi-split air conditioners, characterized in that, Applied to an outdoor unit, wherein the outdoor unit is connected to multiple indoor units, the method includes: Receive the time-series feature vectors uploaded by each indoor unit, and determine the global heat load prediction value based on the time-series feature vectors of each indoor unit; Based on the global heat load forecast, a recommended refrigerant flow rate and recommended fan speed are generated for each indoor unit; Collect local decision broadcast frames from all indoor units, wherein each local decision broadcast frame contains the comfort deviation and current fan speed of each indoor unit; Based on the global heat load forecast, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate level, and the recommended fan speed, global optimization is performed with the goal of optimal overall energy efficiency under the condition of satisfying refrigerant constraints. Recommended control commands are generated for the indoor units to be optimized and sent to the corresponding indoor units.

2. The coordinated control method for multi-split air conditioners according to claim 1, characterized in that, The outdoor unit and the indoor unit are each deployed with a pre-trained air conditioning heat load prediction model, which includes a feature extraction network and a load prediction network; the time-series feature vector is extracted from the original sensor data by the indoor unit based on the feature extraction network of the air conditioning heat load prediction model deployed on itself; The step of determining the global heat load prediction value based on the time-series feature vectors of each indoor unit includes: The time-series feature vectors of each indoor unit are aggregated and then input into the load prediction network, which outputs the global heat load prediction value.

3. The coordinated control method for multi-split air conditioners according to claim 1, characterized in that, The step of generating recommended refrigerant flow rates and recommended fan speeds for each indoor unit based on the global heat load forecast includes: The total refrigerant demand and total air volume demand are determined based on the global heat load forecast. Obtain the rated capacity or historical load percentage of each indoor unit, and allocate recommended refrigerant flow to each indoor unit according to the rated capacity or historical load percentage of each indoor unit; the sum of the recommended refrigerant flow of all indoor units is the total refrigerant requirement. Based on the total air volume requirement and the rated air volume of each indoor unit, determine the fan speed of each indoor unit.

4. The coordinated control method for multi-split air conditioners according to claim 1, characterized in that, The local decision broadcast frame also includes the local computing power margin of the corresponding indoor unit; the global optimization is performed based on the global heat load prediction value, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate level, and the recommended fan speed, with the goal of optimal comprehensive energy efficiency under the condition of satisfying refrigerant constraints, generating recommended control instructions for the indoor unit to be optimized, and sending the recommended control instructions to the corresponding indoor unit, including: Determine whether the comfort deviation of each indoor unit is within the preset deviation range, and identify indoor units whose comfort deviation is not within the preset deviation range as indoor units to be optimized; and / or, perform a computing power load balancing check based on the local computing power margin of each indoor unit, and identify indoor units whose load rate deviates from the average load rate by more than the preset load rate deviation threshold as indoor units to be optimized. Based on the global heat load forecast, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate level, and the recommended fan speed, global optimization is performed with the goal of optimal overall energy efficiency under the condition of satisfying refrigerant constraints. Recommended control commands are generated for the indoor units to be optimized and sent to the corresponding indoor units to be optimized.

5. The coordinated control method for multi-split air conditioners according to claim 4, characterized in that, The outdoor unit is equipped with a global scheduling model; the refrigerant constraints include compressor frequency being within the upper and lower limits, total refrigerant flow being conserved, evaporator outlet superheat not being lower than a preset value, and computing power load balancing. Based on the global heat load forecast, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate, and the recommended fan speed, global optimization is performed with the goal of achieving optimal overall energy efficiency under refrigerant constraints. This generates recommended control commands for the indoor unit to be optimized, including: Construct a state vector, which includes the load rate, comfort deviation, and current fan speed of the indoor unit to be optimized, as well as the compressor frequency, condensing pressure, outdoor temperature, and current electricity price of the outdoor unit; The state vector, the global heat load prediction value, the recommended refrigerant flow rate level, and the recommended fan speed are input into the global scheduling model. Under the condition of satisfying refrigerant constraints, with the goal of optimal comprehensive energy efficiency, the global scheduling model outputs parameter adjustment instructions and computing power scheduling instructions for the indoor unit to be optimized. The parameter adjustment instructions are used to instruct the opening degree of the electronic expansion valve of the indoor unit to be optimized to be set to the target opening degree and the compressor frequency to be set to the target frequency. The computing power scheduling instructions are used to migrate some inference tasks of the indoor units with load rates higher than the preset load rate threshold to the indoor units with load rates lower than the preset load rate threshold, or to adjust the inference frequency of the corresponding indoor units.

6. The coordinated control method for multi-split air conditioners according to claim 5, characterized in that, The step of outputting parameter adjustment instructions and computing power scheduling instructions for the indoor unit to be optimized through the global scheduling model includes: The global scheduling model uses the comprehensive energy efficiency index as the reward target and simultaneously outputs computing power scheduling actions and cooling control actions. The comprehensive energy efficiency index is the ratio of cooling capacity to comprehensive power consumption, which includes compressor power consumption, the operating power consumption of all chips in the indoor and outdoor units, and the communication bus power consumption. The computing power scheduling actions include executing inference tasks locally, offloading inference tasks to the outdoor unit or other indoor units, and adjusting the frequency of inference tasks executed by the indoor units. The cooling control actions include compressor frequency and electronic expansion valve opening.

7. The coordinated control method for multi-split air conditioners according to claim 1, characterized in that, After outputting the computing power scheduling command for the indoor unit to be optimized, the method further includes: A credit score is maintained for each indoor unit, which is used to assess the reliability of the indoor unit as an offloading target; the offloading target is the indoor unit that receives and executes the inference task of offloading other indoor units. Candidate unloading target sequences are generated by sorting the effective computing power of each indoor unit in descending order, and the candidate unloading target sequences are sent to the indoor units; the effective computing power is calculated based on the indoor unit load rate and the credit score.

8. The coordinated control method for multi-split air conditioners according to claim 7, characterized in that, The method further includes: The credit score is updated based on the execution result of the unloading task when the indoor unit is the unloading target and the heartbeat status of the indoor unit; When the indoor unit's credit score is lower than a preset threshold, the indoor unit will be temporarily suspended from being used as the unloading target, an indoor unit stability alarm will be issued, and the indoor unit will be triggered to perform a self-test. After the indoor unit passes the self-test, the credit score of the indoor unit will be restored to the preset value.

9. The coordinated control method for multi-split air conditioners according to claim 8, characterized in that, The execution result of the unloading task includes whether the inference task of unloading other indoor units by the indoor unit was successfully completed, the return time, and whether the return result verification passed; the updating of the credit score based on the execution result of the unloading task when the indoor unit is the unloading target and the heartbeat status of the indoor unit includes: When an indoor unit successfully completes the unloading task of other indoor units and the return time is less than the expected time limit, the credit score is increased; When a task execution error, heartbeat error, or indoor unit actively refuses to uninstall the request occurs, the credit score is deducted; the task execution error includes failure to return a verification result; the heartbeat error includes the indoor unit failing to send a computing power heartbeat frame to the outdoor unit for a preset number of consecutive times.

10. A collaborative control method for multi-split air conditioners, characterized in that, Applied to an indoor unit, wherein the indoor unit is connected to an outdoor unit and other indoor units respectively, the method includes: The collected environmental data is used to extract features to obtain a time-series feature vector, and the time-series feature vector is uploaded to the outdoor unit; The comfort deviation is calculated based on the environmental data and the temporal feature vector. Obtain the current fan speed, and encapsulate the comfort deviation and the current fan speed into a local decision broadcast frame for broadcasting; Receive the recommended control command issued by the outdoor unit; The system outputs local control commands based on the current operating status of the indoor unit, arbitrates the recommended control commands with the local control commands, generates target control commands based on the arbitration result, and executes the target control commands.

11. The coordinated control method for multi-split air conditioners according to claim 10, characterized in that, The outdoor unit and the indoor unit are each deployed with a pre-trained air conditioning heat load prediction model, which includes a feature extraction network and a load prediction network. The global heat load prediction value is obtained by the outdoor unit based on the load prediction network of the air conditioning heat load prediction model deployed on it, based on the time-series feature vector. The step of extracting time-series feature vectors from the collected environmental data includes: The collected environmental data is input into the feature extraction network, and the time-series feature vector is output through the feature extraction network.

12. The coordinated control method for multi-split air conditioners according to claim 10, characterized in that, The step of calculating the comfort deviation based on the environmental data and the time-series feature vector includes: Based on the environmental data, temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value are determined; the human body heat dissipation contribution value characterizes the contribution of heat emitted by people in the room to the room's heat load. The weighting coefficients corresponding to the temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value are determined based on the time-series feature vector. The comfort deviation is obtained by weighting and synthesizing the temperature deviation, humidity deviation, carbon dioxide concentration deviation, and human body heat dissipation contribution value with corresponding weighting coefficients.

13. The coordinated control method for multi-split air conditioners according to claim 10, characterized in that, The step of arbitrating the recommended control command with the local control command, and generating a target control command based on the arbitration result, includes: Obtain the first confidence level corresponding to the recommended control instruction and the second confidence level corresponding to the local control instruction, and calculate the deviation between the corresponding control quantities in the recommended control instruction and the local control instruction; If the first confidence level is higher than the first confidence level threshold, the second confidence level is higher than the first confidence level threshold, and the deviation is less than the first deviation threshold, then the first confidence level and the second confidence level are used as weights to perform a weighted average of the corresponding control quantities in the recommended control command and the local control command to obtain the target control command. If either the first confidence level or the second confidence level is lower than the second confidence level threshold, then the local control instruction is used as the target control instruction; the first confidence level threshold is greater than the second confidence level threshold. If the deviation is greater than the second deviation threshold, then the local control command is used as the target control command; the first deviation threshold is less than the second deviation threshold.

14. The coordinated control method for multi-split air conditioners according to claim 10, characterized in that, The method further includes: Obtain the current local load rate and the real-time requirements of the inference tasks to be executed; If the current local load rate is lower than a preset load rate threshold and the required completion time of the real-time requirement of the inference task to be executed is less than a preset time threshold, then the inference task to be executed is executed locally. If the current local load rate is greater than or equal to a preset load rate threshold, or if the inference task to be executed needs to integrate data from other indoor units, the inference task to be executed will be sent to the outdoor unit or other indoor units with load rates lower than the preset load rate threshold via the offload protocol.

15. The coordinated control method for multi-split air conditioners according to claim 14, characterized in that, Sending the inference task to be executed to the outdoor unit or the indoor unit with a load rate lower than a preset load rate threshold via the offloading protocol includes: Obtain the candidate unloading target sequence sent by the outdoor unit; the candidate unloading target sequence is generated by the outdoor unit sorting the outdoor unit in descending order according to the effective computing power, and the effective computing power is calculated based on the indoor unit load rate and credit score; The inference task to be executed is an unload request frame, which is sent to the corresponding indoor unit in sequence according to the candidate unload target sequence; the unload request frame includes task type, input data summary, expected completion time limit and priority level. If the current indoor unit experiences a load increase and returns to a busy indication during inference, the next candidate indoor unit is determined along the candidate unloading target sequence; If all candidate indoor units in the candidate unloading target sequence return a busy indication or time out without response, then fall back to local degradation inference.

16. The coordinated control method for multi-split air conditioners according to claim 10, characterized in that, The method further includes: If the loss of the master node heartbeat of the outdoor unit is detected, it is determined that the communication between the indoor unit and the outdoor unit is interrupted, and the local control command is executed.

17. A collaborative control system for a multi-split air conditioner, characterized in that, The system includes an outdoor unit and multiple indoor units, with each indoor unit connected to the outdoor unit and other indoor units respectively. The indoor unit is used to extract features from the collected environmental data to obtain a time-series feature vector, and then upload the time-series feature vector to the outdoor unit. The outdoor unit is used to receive the time-series feature vectors uploaded by each indoor unit and determine the global heat load prediction value based on the time-series feature vectors of each indoor unit. The indoor unit is used to calculate the comfort deviation based on the environmental data and the time-series feature vector; obtain the current fan speed; and encapsulate the comfort deviation and the current fan speed into a local decision broadcast frame for broadcasting. The outdoor unit is used to generate recommended refrigerant flow rate and recommended fan speed for each indoor unit based on the global heat load prediction value; collect local decision broadcast frames broadcast by all indoor units; and perform global optimization with the goal of optimal comprehensive energy efficiency under the condition of satisfying refrigerant constraints, based on the global heat load prediction value, the comfort deviation of each indoor unit, the current fan speed, the recommended refrigerant flow rate and the recommended fan speed, to generate recommended control commands for each indoor unit and send the recommended control commands to the corresponding indoor units. The indoor unit is used to receive recommended control commands from the outdoor unit and output local control commands according to the current operating status of the indoor unit. The recommended control command is arbitrated with the local control command, and a target control command is generated based on the arbitration result and executed.

18. A multi-split air conditioner, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the coordinated control method for a multi-split air conditioner as claimed in any one of claims 1-9 and 10-16.

19. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the coordinated control method for a multi-split air conditioner as described in any one of claims 1-9 and 10-16.

Images