Air conditioner temperature non-sensitive control method based on online learning
By using multi-source reliability gating and continuous optimization mechanism for switching penalties, the problems of predictive instability and frequent equipment adjustments in air conditioning control are solved, achieving stable air conditioning operation and consistent user experience, while reducing energy consumption and equipment start-up and shutdown frequency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-29
- Publication Date
- 2026-04-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing air conditioning control methods struggle to achieve stable prediction, continuous optimization, and adaptive adjustment under multi-source uncertain input conditions, resulting in high energy consumption or frequent fluctuations in indoor temperature, and lacking smooth control over user comfort.
By employing a Gaussian process state-space model with multi-source reliability gating and a continuous optimization mechanism with switching penalty, the system evaluates the multi-source operating parameters of the air conditioner and iteratively solves the control variables through online learning. This constructs an integrated, seamless adjustment process that combines prediction, optimization, control, and learning, achieving continuity of cooling and heating load states and stability of user experience.
Maintaining the continuity and consistency of load forecasting results under complex operating conditions, reducing ineffective switching behavior during equipment operation, and achieving stable regulation and consistent user experience control during long-term air conditioning operation.
Smart Images
Figure CN121855028A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent building environment control and HVAC control technology, and in particular to a non-intrusive air conditioning temperature control method based on online learning. Background Technology
[0002] With the continuous expansion of building scale and the sustained rise in energy costs, HVAC accounts for an increasingly larger proportion of building energy consumption. During air conditioning operation, a trade-off must usually be struck between energy consumption control and indoor thermal environment stability. Traditional air conditioning control methods often employ fixed setpoint control, zoned time-of-day control, or adjustment strategies based on empirical rules. These methods rely on manually set parameters and are difficult to adapt to dynamic fluctuations in external weather conditions, indoor load changes, and equipment operating status, easily leading to problems such as high energy consumption or frequent fluctuations in indoor temperature.
[0003] In recent years, with the development of sensing technology and computing power, some air conditioning controls have begun to introduce predictive models or optimization control methods to estimate load changes in advance and adjust operating strategies accordingly. However, existing predictive models in practical applications typically assume that multi-source input data has consistent reliability. When sensor data is missing, fluctuates abnormally, or there are sudden environmental changes, the prediction results are prone to deviation, leading to frequent corrections of the control strategy. In addition, existing optimization control methods often use discrete rules or threshold judgments when dealing with changes in operating states such as switching between heat and cold sources and equipment start-up and shutdown. These methods are difficult to model uniformly within a continuous optimization framework, which can easily cause equipment instability and energy consumption fluctuations.
[0004] On the other hand, users' perception of the indoor thermal environment is not only related to the temperature setpoint, but also closely related to factors such as the rate of temperature change and the frequency of operating state switching. Existing control methods often focus on energy consumption or setpoint temperature control, lacking comprehensive constraints on the continuity of operation and perceived stability, making it difficult to achieve smooth adjustment of user comfort while ensuring energy efficiency. Therefore, how to achieve stable prediction, continuous optimization, and adaptive adjustment of air conditioning under multi-source uncertain input conditions remains a problem that urgently needs to be solved in existing technologies. Summary of the Invention
[0005] One objective of this invention is to propose a non-intrusive temperature control method for air conditioners based on online learning. This invention introduces a state-space prediction model with multi-source reliability gating and a continuous optimization mechanism for switching penalties. It performs online evaluation, load prediction, and iterative solution of control variables for multi-source operating parameters of the air conditioner, and constructs a non-intrusive adjustment process integrating prediction, optimization, control, and learning. It has the advantages of stable energy consumption control, controlled switching of operating states, and stable user experience.
[0006] According to an embodiment of the present invention, an air conditioning temperature non-intrusive control method based on online learning includes the following steps:
[0007] S1. Collect indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters to form a multi-source parameter time series. S2. Construct a Gaussian process state-space model incorporating a multi-source reliability gating factor. Adjust the weights of various parameters in the state transition calculation using the multi-source reliability gating factor to generate a predicted sequence of cooling and heating load states within the future time window. S3. Construct a nonlinear optimization problem based on the predicted cooling and heating load states, including cooling and heating source state variables, equipment continuous operation variables, and constraints. S4. Set a switching penalty continuous term in the nonlinear optimization problem. The switching penalty continuous term represents changes in cooling and heating source operating modes and equipment start-up and shutdown changes. S5. Use a sequential quadratic programming algorithm to iteratively solve the nonlinear optimization problem containing the switching penalty continuous term, generating a control variable update sequence. S6. Generate an air conditioning operation control strategy based on the control variable update sequence. The operation control strategy includes cooling and heating source operating modes, equipment operating frequencies, and set parameters. S7. Calculate the prediction deviation after the control cycle ends. Use the prediction deviation to update the multi-source reliability gating factor value, completing the online learning process.
[0008] Optionally, S2 specifically includes:
[0009] S21. Read indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters from the multi-source parameter time series in time index order. Perform numerical normalization on each type of parameter, and the normalization result forms a multi-source state input vector sequence. S22. According to the parameter category classification rules of the multi-source state input vector sequence, assign a unique reliability identifier to each type of parameter. Initialize the multi-source reliability gating factor vector based on the reliability identifier. Each component value in the multi-source reliability gating factor vector corresponds one-to-one with each type of parameter. S23. Construct a Gaussian process state-space model, defining the state vector as containing cold and heat load state components and the observation vector as containing the multi-source state input vector sequence. Configure the state transition function in the Gaussian process state-space model, expressing the state transition function in kernel function form. S24. Introduce the multi-source reliability gating factor into the kernel function calculation structure. The vector, with the corresponding components in the multi-source reliability gating factor vector as weight coefficients, is applied to the similarity calculation process of the corresponding parameter components in the multi-source state input vector sequence, forming a state transition kernel function with gating weights; S25, the state transition calculation is performed on the state transition kernel function with gating weights to obtain the state transition results of cold and hot loads between adjacent time index positions, and the state transition results are written into the state prediction buffer; S26, according to the preset prediction step size, the cold and hot load state values of future consecutive time index positions are recursively calculated from the state prediction buffer, and the state transition kernel function with gating weights is called at each time index position in the recursive calculation process; S27, the cold and hot load state values obtained at each time index position are arranged in chronological order to form a cold and hot load state prediction sequence within the future time window, and the cold and hot load state prediction sequence is output to step S3.
[0010] Optionally, the multi-source reliability gating factor vector specifically includes:
[0011] Extract time segments from multi-source parameter time series that cover the current time index position and the consecutive time index positions preceding it;
[0012] Indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters are read within a time segment, and corresponding parameter subsequences are formed according to the parameter category classification rules;
[0013] Perform missing value labeling and statistical operations on each parameter subsequence to obtain the missing value count and subsequence length;
[0014] Perform a value range verification operation on each parameter subsequence, compare the parameter value with the preset lower limit and upper limit of the parameter value by time index, and accumulate the out-of-bounds count value;
[0015] Perform adjacent difference operations on each parameter subsequence, calculate the difference values of adjacent time index positions in time index order, and form a difference sequence;
[0016] Perform mutation counting operation on the difference sequence, compare the absolute value of the difference value with the preset difference threshold by time indexing, and accumulate the over-threshold count value;
[0017] Perform statistical operations on each parameter subsequence, calculate the subsequence mean and subsequence variance, and write the mean and variance into the reliability calculation buffer;
[0018] For each parameter category, construct a reliability intermediate vector component. The reliability intermediate vector component is formed by arranging missing counts, out-of-bounds counts, over-threshold counts, and variance values in a preset order.
[0019] Normalization mapping is performed on the reliability intermediate vector components. Linear combination calculation is performed by mapping the weight vector and mapping the bias scalar. The calculation results are truncated in the range of 0 to 1 to generate the reliability component values.
[0020] Arrange the reliability component values of each parameter in a fixed order according to the parameter category to form a multi-source reliability gating factor vector, and write the multi-source reliability gating factor vector into the gating factor cache area.
[0021] Optionally, S3 specifically includes:
[0022] Read the cold and heat load state values corresponding to each predicted time position in the time index order from the cold and heat load state prediction sequence; construct an optimization variable set, which includes cold and heat source operating state variables and equipment continuous operation variables; construct an objective function, which includes energy consumption cost terms and comfort constraint cost terms; construct a constraint condition set, which includes cold and heat source operating state constraints, equipment continuous operation constraints, and operating boundary constraints; combine the objective function, optimization variable set, and constraint condition set to form a nonlinear optimization problem, and write the nonlinear optimization problem into the optimization solution buffer.
[0023] Optionally, S4 includes the following steps:
[0024] The system reads the operating state variables of the cold and heat sources and the continuous operating variables of the equipment from the nonlinear optimization problem, and constructs a state change sequence of adjacent time index positions according to the time index order. It then performs differential calculations on the state change sequences, which include calculating the changes in the operating state variables of the cold and heat sources and the changes in the continuous operating variables of the equipment at adjacent time index positions. Next, it performs amplitude mapping operations on the changes, mapping them to numerical values within a continuous range. Finally, it constructs a switching penalty function, which takes the numerical value of the change as input and outputs the corresponding continuous penalty value. These continuous penalty values are then accumulated in time index order to form a continuous switching penalty term. This continuous switching penalty term is then written as an independent cost term into the objective function of the nonlinear optimization problem.
[0025] Optionally, S5 specifically includes:
[0026] Read the set of optimization variables, the set of constraints, the objective function, and the switching penalty continuous term in the nonlinear optimization problem, and establish the iterative initialization structure for the sequential quadratic programming algorithm;
[0027] In the iterative initialization structure, write the initial value vectors of the cold and heat source operating state variables, the initial value vectors of the equipment continuous operation variables, and the initial value vectors of the set parameters. The initial value vectors are arranged in time index order to form the control variable initial value sequence.
[0028] In the current iteration, gradient calculation is performed on the objective function at the initial value sequence of control variables, and Jacobian matrix calculation is performed on the constraint set at the initial value sequence of control variables. The gradient calculation results and Jacobian matrix calculation results are written into the quadratic approximation buffer.
[0029] In the current iteration, the switching penalty continuous term is calculated by differential chaining at the initial value sequence of the control variables to obtain the switching penalty gradient sequence. The switching penalty gradient sequence and the objective function gradient are written into the quadratic approximation buffer.
[0030] A quadratic approximation objective function is constructed in the quadratic approximation buffer. The quadratic approximation objective function consists of the objective function gradient, the switching penalty gradient sequence, and the quadratic approximation Hessian matrix.
[0031] Linearization constraints are constructed in the quadratic approximation buffer. The linearization constraints consist of the Jacobian matrix of the constraint condition set and the constraint residual vector.
[0032] Solve the quadratic programming subproblem by applying the quadratic approximation objective function and linearization constraints, and output the incremental sequence of control variables.
[0033] Perform a step scaling operation on the control variable increment sequence, and the step scaling operation outputs the step scaling coefficient to generate a scaled increment sequence;
[0034] Update the initial value sequence of control variables using the scaling increment sequence, and write the update result into the initial value sequence of control variables for the next iteration.
[0035] After each iteration, the change in the objective function value and the norm of the constraint residual vector are calculated. When the termination threshold is met, the final control variable update sequence is output and written into the control variable update sequence buffer.
[0036] Optionally, S6 includes the following steps:
[0037] The control variable update vectors are read from the control variable update sequence buffer in chronological order. A vector summation operation is performed between the control variable update vectors and the initial control variable value vectors at the corresponding time indices to generate a control variable result vector. A variable type parsing operation is performed on the control variable result vector, splitting it into sub-vectors of cold / heat source operating status variables, equipment continuous operation variables, and set parameter sub-vectors. A state mapping operation is performed on the cold / heat source operating status variable sub-vectors to generate cold / heat source operating mode identifiers. An amplitude constraint operation is performed on the equipment continuous operation variable sub-vectors to form an equipment operating frequency sequence. A parameter parsing operation is performed on the set parameter sub-vectors to obtain set parameter values. The cold / heat source operating mode identifiers, equipment operating frequency sequences, and set parameter values are combined in chronological order to form the air conditioning operation control strategy, which is then written into the operation control strategy buffer.
[0038] Optionally, S7 specifically includes:
[0039] At the end of the control cycle, the control cycle termination time index is located from the multi-source parameter time series. The indoor environmental parameters and equipment operating parameters at the corresponding time index are read, and the actual cooling and heating load status values are obtained according to the parameter position parsing rules. The predicted cooling and heating load status values corresponding to the termination time index are read from the cooling and heating load status prediction sequence. The difference operation is performed between the predicted cooling and heating load status values and the actual cooling and heating load status values to generate the prediction deviation value. If the prediction sequence covers several time index positions, the predicted cooling and heating load status values before the termination time index position and the actual cooling and heating load status values at the corresponding time index positions are read. The difference operation is performed on each time index, and the values are arranged in chronological order to form the prediction deviation sequence. The amplitude mapping operation is performed on the prediction deviation sequence to generate the deviation measurement sequence. The deviation measurement sequence is written into the prediction deviation buffer as input data for the multi-source reliability gating factor update operation.
[0040] The beneficial effects of this invention are:
[0041] (1) This invention introduces a Gaussian process state-space model with multi-source reliability gating factors to quantitatively model the stability of indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters and energy price parameters. In the state transition calculation process, the participation weights of different parameters are dynamically adjusted so that the cold and heat load prediction process can effectively suppress the influence of abnormal data and unstable inputs on the prediction results. This maintains the continuity and consistency of load prediction results under complex operating conditions and provides a stable data basis for subsequent control decisions.
[0042] (2) In the process of nonlinear optimization modeling, the present invention introduces a continuous switching penalty term, which maps the changes in the operating mode of the cold and heat source and the start-up and shutdown behavior of the equipment into a continuous cost component. Combined with the sequential quadratic programming algorithm, the control variables are iteratively solved, so that the update process of the control variables tends to evolve smoothly in mathematical structure, avoiding frequent large-scale adjustments to the operating state, thereby reducing the ineffective switching behavior in the equipment operation process and maintaining the continuity and controllability of the air conditioning operating state.
[0043] (3) This invention constructs an online learning closed loop between the prediction model and the control strategy by calculating the prediction deviation after the control cycle ends and feeding it back to the multi-source reliability gating factor update process. This enables the prediction weight allocation and control decision basis to be automatically adjusted according to changes in the operating environment and load characteristics. It can continuously adapt to changes in operating conditions without human intervention, and achieve stable regulation and consistent control of user experience during long-term operation of the air conditioner. Attached Figure Description
[0044] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0045] Figure 1 This is a flowchart of a non-intrusive air conditioning temperature control method based on online learning proposed in this invention.
[0046] Figure 2 This diagram illustrates the implementation of a multi-source reliability gating prediction method for non-intrusive air conditioning temperature control based on online learning, as proposed in this invention.
[0047] Figure 3 This invention presents a continuous optimization control graph for switching penalty in a non-intrusive air conditioning temperature control method based on online learning. Detailed Implementation
[0048] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0049] refer to Figure 1-3 A non-intrusive temperature control method for air conditioners based on online learning includes the following steps:
[0050] S1. Collect indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters to form a multi-source parameter time series. S2. Construct a Gaussian process state-space model incorporating a multi-source reliability gating factor. Adjust the weights of various parameters in the state transition calculation using the multi-source reliability gating factor to generate a predicted sequence of cooling and heating load states within the future time window. S3. Construct a nonlinear optimization problem based on the predicted cooling and heating load states, including cooling and heating source state variables, equipment continuous operation variables, and constraints. S4. Set a switching penalty continuous term in the nonlinear optimization problem. The switching penalty continuous term represents changes in cooling and heating source operating modes and equipment start-up and shutdown changes. S5. Use a sequential quadratic programming algorithm to iteratively solve the nonlinear optimization problem containing the switching penalty continuous term, generating a control variable update sequence. S6. Generate an air conditioning operation control strategy based on the control variable update sequence. The operation control strategy includes cooling and heating source operating modes, equipment operating frequencies, and set parameters. S7. Calculate the prediction deviation after the control cycle ends. Use the prediction deviation to update the multi-source reliability gating factor value, completing the online learning process.
[0051] In this embodiment, S2 specifically includes:
[0052] S21. Read indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters from the multi-source parameter time series in time index order. Perform numerical normalization on each type of parameter, and the normalization result forms a multi-source state input vector sequence. S22. According to the parameter category classification rules of the multi-source state input vector sequence, assign a unique reliability identifier to each type of parameter. Initialize the multi-source reliability gating factor vector based on the reliability identifier. Each component value in the multi-source reliability gating factor vector corresponds one-to-one with each type of parameter. S23. Construct a Gaussian process state-space model, defining the state vector as containing cold and heat load state components and the observation vector as containing the multi-source state input vector sequence. Configure the state transition function in the Gaussian process state-space model, expressing the state transition function in kernel function form. S24. Introduce the multi-source reliability gating factor into the kernel function calculation structure. The vector, with the corresponding components in the multi-source reliability gating factor vector as weight coefficients, is applied to the similarity calculation process of the corresponding parameter components in the multi-source state input vector sequence, forming a state transition kernel function with gating weights; S25, the state transition calculation is performed on the state transition kernel function with gating weights to obtain the state transition results of cold and hot loads between adjacent time index positions, and the state transition results are written into the state prediction buffer; S26, according to the preset prediction step size, the cold and hot load state values of future consecutive time index positions are recursively calculated from the state prediction buffer, and the state transition kernel function with gating weights is called at each time index position in the recursive calculation process; S27, the cold and hot load state values obtained at each time index position are arranged in chronological order to form a cold and hot load state prediction sequence within the future time window, and the cold and hot load state prediction sequence is output to step S3.
[0053] In this embodiment, the multi-source reliability gating factor vector specifically includes:
[0054] Extract time segments from multi-source parameter time series that cover the current time index position and the consecutive time index positions preceding it;
[0055] Indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters are read within a time segment, and corresponding parameter subsequences are formed according to the parameter category classification rules;
[0056] Perform missing value labeling and statistical operations on each parameter subsequence to obtain the missing value count and subsequence length;
[0057] Perform a value range verification operation on each parameter subsequence, compare the parameter value with the preset lower limit and upper limit of the parameter value by time index, and accumulate the out-of-bounds count value;
[0058] Perform adjacent difference operations on each parameter subsequence, calculate the difference values of adjacent time index positions in time index order, and form a difference sequence;
[0059] Perform mutation counting operation on the difference sequence, compare the absolute value of the difference value with the preset difference threshold by time indexing, and accumulate the over-threshold count value;
[0060] Perform statistical operations on each parameter subsequence, calculate the subsequence mean and subsequence variance, and write the mean and variance into the reliability calculation buffer;
[0061] For each parameter category, construct a reliability intermediate vector component. The reliability intermediate vector component is formed by arranging missing counts, out-of-bounds counts, over-threshold counts, and variance values in a preset order.
[0062] Normalization mapping is performed on the reliability intermediate vector components. Linear combination calculation is performed by mapping the weight vector and mapping the bias scalar. The calculation results are truncated in the range of 0 to 1 to generate the reliability component values.
[0063] Arrange the reliability component values of each parameter in a fixed order according to the parameter category to form a multi-source reliability gating factor vector, and write the multi-source reliability gating factor vector into the gating factor cache area.
[0064] In this embodiment, S3 specifically includes:
[0065] Read the cold and heat load state values corresponding to each predicted time position in the time index order from the cold and heat load state prediction sequence; construct an optimization variable set, which includes cold and heat source operating state variables and equipment continuous operation variables; construct an objective function, which includes energy consumption cost terms and comfort constraint cost terms; construct a constraint condition set, which includes cold and heat source operating state constraints, equipment continuous operation constraints, and operating boundary constraints; combine the objective function, optimization variable set, and constraint condition set to form a nonlinear optimization problem, and write the nonlinear optimization problem into the optimization solution buffer.
[0066] In this embodiment, step S4 includes the following steps:
[0067] The system reads the operating state variables of the cold and heat sources and the continuous operating variables of the equipment from the nonlinear optimization problem, and constructs a state change sequence of adjacent time index positions according to the time index order. It then performs differential calculations on the state change sequences, which include calculating the changes in the operating state variables of the cold and heat sources and the changes in the continuous operating variables of the equipment at adjacent time index positions. Next, it performs amplitude mapping operations on the changes, mapping them to numerical values within a continuous range. Finally, it constructs a switching penalty function, which takes the numerical value of the change as input and outputs the corresponding continuous penalty value. These continuous penalty values are then accumulated in time index order to form a continuous switching penalty term. This continuous switching penalty term is then written as an independent cost term into the objective function of the nonlinear optimization problem.
[0068] In this embodiment, the penalty function is specifically:
[0069] This is a mathematical mapping structure that maps changes in operating state to continuous penalty values, used to characterize the intensity of changes in the operating state variables of cold and heat sources and the continuous operating variables of equipment between adjacent time index positions. The function takes the calculated change amplitude values from the state change sequence as input and outputs the corresponding penalty values through a preset continuous mapping relationship, thereby transforming the originally discrete changes in operating modes and equipment start-up and shutdown behaviors into cost components that can participate in continuous optimization solutions.
[0070] The switching penalty function employs a continuous monotonic mapping structure to perform numerical mapping processing on the input change magnitude. When the change magnitude is close to zero, the penalty value output by the switching penalty function is close to zero, indicating that the running state remains basically continuous between adjacent time index positions. As the change magnitude increases, the penalty value output by the switching penalty function increases continuously, reflecting the cumulative impact of the intensity of the change in running state.
[0071] This mapping process does not rely on discrete threshold judgments, but distinguishes different levels of change through continuous functional relationships, so that any tiny change corresponds to a unique penalty value.
[0072] In this embodiment, S5 specifically includes:
[0073] Read the set of optimization variables, the set of constraints, the objective function, and the switching penalty continuous term in the nonlinear optimization problem, and establish the iterative initialization structure for the sequential quadratic programming algorithm;
[0074] In the iterative initialization structure, write the initial value vectors of the cold and heat source operating state variables, the initial value vectors of the equipment continuous operation variables, and the initial value vectors of the set parameters. The initial value vectors are arranged in time index order to form the control variable initial value sequence.
[0075] In the current iteration, gradient calculation is performed on the objective function at the initial value sequence of control variables, and Jacobian matrix calculation is performed on the constraint set at the initial value sequence of control variables. The gradient calculation results and Jacobian matrix calculation results are written into the quadratic approximation buffer.
[0076] In the current iteration, the switching penalty continuous term is calculated by differential chaining at the initial value sequence of the control variables to obtain the switching penalty gradient sequence. The switching penalty gradient sequence and the objective function gradient are written into the quadratic approximation buffer.
[0077] A quadratic approximation objective function is constructed in the quadratic approximation buffer. The quadratic approximation objective function consists of the objective function gradient, the switching penalty gradient sequence, and the quadratic approximation Hessian matrix.
[0078] Linearization constraints are constructed in the quadratic approximation buffer. The linearization constraints consist of the Jacobian matrix of the constraint condition set and the constraint residual vector.
[0079] Solve the quadratic programming subproblem by applying the quadratic approximation objective function and linearization constraints, and output the incremental sequence of control variables.
[0080] Perform a step scaling operation on the control variable increment sequence, and the step scaling operation outputs the step scaling coefficient to generate a scaled increment sequence;
[0081] Update the initial value sequence of control variables using the scaling increment sequence, and write the update result into the initial value sequence of control variables for the next iteration.
[0082] After each iteration, the change in the objective function value and the norm of the constraint residual vector are calculated. When the termination threshold is met, the final control variable update sequence is output and written into the control variable update sequence buffer.
[0083] In this embodiment, the second-order approximate Hessian matrix specifically includes:
[0084] At the beginning of each iteration of the sequential quadratic programming algorithm, the gradient of the objective function is first calculated at the current initial value sequence of the control variables. The gradient of the energy cost term is then merged with the gradient of the switching penalty continuity term to form a complete objective function gradient vector. This gradient vector serves as the first-order information input for the current iteration and is written into the quadratic approximation buffer.
[0085] Subsequently, the changes in the initial value sequence of the control variables and the changes in the gradient vector of the objective function are recorded between adjacent iterations. By comparing the correspondence between the gradient change direction and the variable change direction in adjacent iterations, a numerical update structure for characterizing the local curvature change trend is constructed. This update structure does not rely on the explicit calculation of the second-order partial derivative of the objective function, but rather approximates the curvature information through the numerical relationship of the first-order information change.
[0086] During the formation process, numerical consistency processing is performed on the updated matrix structure to ensure that its dimensions are consistent with the sequence of control variables, and the symmetry of the matrix is corrected to guarantee that it can participate in subsequent calculations as a second-order term in the quadratic approximation objective function. Where necessary, magnitude constraints are applied to the values in specific directions of the matrix to avoid local numerical anomalies affecting the stability of the optimization solution.
[0087] Finally, the matrix formed after the above numerical update and correction process is used as the second-order approximate Hessian matrix, which is written into the second-order approximation buffer and together with the objective function gradient vector and the linearized constraint structure, constitutes the second-order approximate objective function for solving the quadratic programming subproblem.
[0088] In this embodiment, step S6 includes the following steps:
[0089] The control variable update vectors are read from the control variable update sequence buffer in chronological order. A vector summation operation is performed between the control variable update vectors and the initial control variable value vectors at the corresponding time indices to generate a control variable result vector. A variable type parsing operation is performed on the control variable result vector, splitting it into sub-vectors of cold / heat source operating status variables, equipment continuous operation variables, and set parameter sub-vectors. A state mapping operation is performed on the cold / heat source operating status variable sub-vectors to generate cold / heat source operating mode identifiers. An amplitude constraint operation is performed on the equipment continuous operation variable sub-vectors to form an equipment operating frequency sequence. A parameter parsing operation is performed on the set parameter sub-vectors to obtain set parameter values. The cold / heat source operating mode identifiers, equipment operating frequency sequences, and set parameter values are combined in chronological order to form the air conditioning operation control strategy, which is then written into the operation control strategy buffer.
[0090] In this embodiment, S7 specifically includes:
[0091] At the end of the control cycle, the control cycle termination time index is located from the multi-source parameter time series. The indoor environmental parameters and equipment operating parameters at the corresponding time index are read, and the actual cooling and heating load status values are obtained according to the parameter position parsing rules. The predicted cooling and heating load status values corresponding to the termination time index are read from the cooling and heating load status prediction sequence. The difference operation is performed between the predicted cooling and heating load status values and the actual cooling and heating load status values to generate the prediction deviation value. If the prediction sequence covers several time index positions, the predicted cooling and heating load status values before the termination time index position and the actual cooling and heating load status values at the corresponding time index positions are read. The difference operation is performed on each time index, and the values are arranged in chronological order to form the prediction deviation sequence. The amplitude mapping operation is performed on the prediction deviation sequence to generate the deviation measurement sequence. The deviation measurement sequence is written into the prediction deviation buffer as input data for the multi-source reliability gating factor update operation.
[0092] Example 1:
[0093] To verify the feasibility and stability of this invention in a real-world operating environment, it was applied to an indoor building environment characterized by the coordinated operation of multiple air conditioning units. This environment is simultaneously affected by external weather changes, indoor occupant activity, and fluctuations in the equipment's own operating status during daily operation. Traditional air conditioning control methods typically rely on fixed setpoints or manual adjustments based on experience, easily leading to problems such as high energy consumption, frequent equipment start-ups and shutdowns, and significant fluctuations in perceived indoor comfort, making it difficult to achieve a balance between energy consumption control and stable comfort. In this application scenario, indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters are continuously collected, forming a continuous multi-source parameter time series. Through unified processing and analysis of these multi-source parameters, the current environmental load change trend and equipment operating status changes can be reflected. For potential abnormal fluctuations, missing data, or instability in the multi-source data during operation, reliability assessments are performed on various parameters, resulting in a multi-source reliability gating factor used to characterize the reliability of different parameters within the current control cycle. In the load forecasting phase, a Gaussian process state-space model incorporating the multi-source reliability gating factor is used to model the heating and cooling load states. By adjusting the weights of multi-source parameters in state transition calculations using a gating factor, the prediction process maintains high sensitivity to stable data and automatically reduces the impact weight of abnormal data, thus obtaining a continuous and smooth heating and cooling load prediction sequence. This prediction result provides a reliable data foundation for subsequent control decisions. In the control decision-making stage, a nonlinear optimization problem is constructed based on the predicted heating and cooling load state sequence, incorporating heating and cooling source operating state variables, continuous equipment operating variables, and operating boundary constraints into the optimization framework. Simultaneously, by introducing a switching penalty continuity term, changes in heating and cooling source operating modes and equipment start-up and shutdown behaviors are mapped as continuous cost components, making the optimization process mathematically inclined to reduce unnecessary state switching. A sequential quadratic programming algorithm is used to iteratively solve the optimization problem, obtaining the control variable update sequence, and generating an air conditioning operation control strategy accordingly. During operation, the equipment is adjusted according to the generated operation control strategy. After the control cycle ends, the predicted heating and cooling load state is compared with the actual heating and cooling load state, the prediction deviation is calculated, and the prediction deviation is fed back to the multi-source reliability gating factor update stage. This online learning mechanism enables gradual adaptation to changes in the environment and equipment characteristics, achieving adaptive adjustment during long-term operation. Statistical analysis of continuous operating data shows that after adopting the method of this invention, the frequency of equipment start-up and shutdown is significantly reduced, the switching of operating states is smoother, the fluctuation range of indoor environmental parameters is reduced, and the overall energy consumption level remains within a relatively stable range. Compared with traditional control methods, this invention effectively suppresses ineffective energy consumption and frequent adjustment behavior while ensuring environmental stability. The table below shows a comparison of key operating indicators under the same operating conditions when using traditional control methods and the control method of this invention.
[0094] Table 1: Comparison of Key Indicators for Air Conditioning Operation
[0095] Comparison indicators Traditional control methods Control method of the present invention Trend of change Stability evaluation Average energy consumption per unit period 100 86 decline Stablize Equipment start-up and shutdown count statistics 42 18 reduce Stablize Number of times the cold / heat source operating mode is switched 36 15 reduce Stablize Indoor temperature fluctuation range 1.8 0.9 Decrease Stablize average prediction deviation 1.2 0.6 Decrease Stablize Number of times the control strategy is adjusted 28 12 reduce Stablize
[0096] Table 1 reflects the differences in several key operating dimensions of air conditioning under the same operating conditions between traditional air conditioning control and the control method of this invention. The average energy consumption per unit cycle shows that under the traditional control method, energy consumption remains consistently high, while under the method of this invention, energy consumption decreases overall and the change is more gradual. This indicates that after the prediction results are incorporated into the control decision, operation no longer relies on frequent adjustments to cope with load fluctuations. The statistical values for the number of equipment start-ups and shutdowns and the number of switching between cold and heat source operating modes are significantly lower under the method of this invention than under the traditional control method. This indicates that by introducing a continuous term for switching penalties, the optimization process actively suppresses unnecessary state changes when updating control variables, resulting in stronger continuity of equipment operating status over time. The indoor temperature fluctuation amplitude shows a significant reduction under the method of this invention, indicating that the control strategy avoids large adjustments to setpoints during execution, which helps maintain stable changes in indoor environmental parameters. This stable change is more consistent with the human body's perception of the thermal environment. The change in the average prediction deviation reflects the consistency of the load forecasting model during long-term operation. Using the method of this invention, the prediction deviation remains at a low level, indicating that the multi-source reliability gating factor can continuously adjust the weights of different parameters in the prediction model, giving the prediction results a strong ability to suppress abnormal data and short-term fluctuations. Furthermore, the number of control strategy adjustments shows that the method of this invention reduces the frequency of strategy updates while ensuring operational stability, avoiding interference from frequent changes in control commands on equipment operation. In summary, the data in the table demonstrate that this invention, through the synergistic effect of prediction, optimization, and online learning, enables the air conditioner to exhibit a more coordinated operating state in terms of energy consumption, operational stability, and continuity of environmental parameters, verifying the feasibility and stability of this invention in practical applications.
[0097] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A non-intrusive temperature control method for air conditioners based on online learning, characterized in that, Includes the following steps: S1. Collect indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters to form a multi-source parameter time series; S2. Construct a Gaussian process state-space model incorporating a multi-source reliability gating factor. Adjust the weights of various parameters in the state transition calculation using the multi-source reliability gating factor to generate a predicted sequence of cooling and heating load states within the future time window. S3. Based on the predicted cooling and heating load states, construct a nonlinear optimization problem including cooling and heating source state variables, equipment continuous operation variables, and constraints. S4. Set a switching penalty continuous term in the nonlinear optimization problem. This term represents changes in cooling and heating source operating modes and equipment start-up and shutdown changes. S5. Use a sequential quadratic programming algorithm to iteratively solve the nonlinear optimization problem including the switching penalty continuous term, generating a control variable update sequence. S6. Generate an air conditioning operation control strategy based on the control variable update sequence. The operation control strategy includes cooling and heating source operating modes, equipment operating frequencies, and set parameters. S7. Calculate the prediction deviation after the control cycle ends. Use the prediction deviation to update the multi-source reliability gating factor values, completing the online learning process.
2. The air conditioning temperature non-intrusive control method based on online learning according to claim 1, characterized in that, S2 specifically includes: S21. Read indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters from the multi-source parameter time series in time index order. Perform numerical normalization on each type of parameter, and the normalization result forms a multi-source state input vector sequence. S22. According to the parameter category classification rules of the multi-source state input vector sequence, assign a unique reliability identifier to each type of parameter. Initialize the multi-source reliability gating factor vector based on the reliability identifier. Each component value in the multi-source reliability gating factor vector corresponds one-to-one with each type of parameter. S23. Construct a Gaussian process state-space model, defining the state vector as containing cold and heat load state components and the observation vector as containing the multi-source state input vector sequence. Configure the state transition function in the Gaussian process state-space model, expressing the state transition function in kernel function form. S24. Introduce the multi-source reliability gating factor into the kernel function calculation structure. The vector, with the corresponding components in the multi-source reliability gating factor vector as weight coefficients, is applied to the similarity calculation process of the corresponding parameter components in the multi-source state input vector sequence, forming a state transition kernel function with gating weights; S25, the state transition calculation is performed on the state transition kernel function with gating weights to obtain the state transition results of cold and hot loads between adjacent time index positions, and the state transition results are written into the state prediction buffer; S26, according to the preset prediction step size, the cold and hot load state values of future consecutive time index positions are recursively calculated from the state prediction buffer, and the state transition kernel function with gating weights is called at each time index position in the recursive calculation process; S27, the cold and hot load state values obtained at each time index position are arranged in chronological order to form a cold and hot load state prediction sequence within the future time window, and the cold and hot load state prediction sequence is output to step S3.
3. The air conditioning temperature non-intrusive control method based on online learning according to claim 2, characterized in that, The multi-source reliability gating factor vector specifically includes: Extract time segments from multi-source parameter time series that cover the current time index position and the consecutive time index positions preceding it; Indoor environmental parameters, outdoor meteorological parameters, equipment operating parameters, and energy price parameters are read within a time segment, and corresponding parameter subsequences are formed according to the parameter category classification rules; Perform missing value labeling and statistical operations on each parameter subsequence to obtain the missing value count and subsequence length; Perform a value range verification operation on each parameter subsequence, compare the parameter value with the preset lower limit and upper limit of the parameter value by time index, and accumulate the out-of-bounds count value; Perform adjacent difference operations on each parameter subsequence, calculate the difference values of adjacent time index positions in time index order, and form a difference sequence; Perform mutation counting operation on the difference sequence, compare the absolute value of the difference value with the preset difference threshold by time indexing, and accumulate the over-threshold count value; Perform statistical operations on each parameter subsequence, calculate the subsequence mean and subsequence variance, and write the mean and variance into the reliability calculation buffer; For each parameter category, construct a reliability intermediate vector component. The reliability intermediate vector component is formed by arranging missing counts, out-of-bounds counts, over-threshold counts, and variance values in a preset order. Normalization mapping is performed on the reliability intermediate vector components. Linear combination calculation is performed by mapping the weight vector and mapping the bias scalar. The calculation results are truncated in the range of 0 to 1 to generate the reliability component values. Arrange the reliability component values of each parameter in a fixed order according to the parameter category to form a multi-source reliability gating factor vector, and write the multi-source reliability gating factor vector into the gating factor cache area.
4. The air conditioning temperature non-intrusive control method based on online learning according to claim 3, characterized in that, S3 specifically includes: Read the cold and heat load state values corresponding to each predicted time position in the time index order from the cold and heat load state prediction sequence; construct an optimization variable set, which includes cold and heat source operating state variables and equipment continuous operation variables; construct an objective function, which includes energy consumption cost terms and comfort constraint cost terms; construct a constraint condition set, which includes cold and heat source operating state constraints, equipment continuous operation constraints, and operating boundary constraints; combine the objective function, optimization variable set, and constraint condition set to form a nonlinear optimization problem, and write the nonlinear optimization problem into the optimization solution buffer.
5. The air conditioning temperature non-sensory control method based on online learning according to claim 4, characterized in that, S4 includes the following steps: The system reads the operating state variables of the cold and heat sources and the continuous operating variables of the equipment from the nonlinear optimization problem, and constructs a state change sequence of adjacent time index positions according to the time index order. It then performs differential calculations on the state change sequences, which include calculating the changes in the operating state variables of the cold and heat sources and the changes in the continuous operating variables of the equipment at adjacent time index positions. Next, it performs amplitude mapping operations on the changes, mapping them to numerical values within a continuous range. Finally, it constructs a switching penalty function, which takes the numerical value of the change as input and outputs the corresponding continuous penalty value. These continuous penalty values are then accumulated in time index order to form a continuous switching penalty term. This continuous switching penalty term is then written as an independent cost term into the objective function of the nonlinear optimization problem.
6. The air conditioning temperature non-sensory control method based on online learning according to claim 5, characterized in that, S5 specifically includes: Read the set of optimization variables, the set of constraints, the objective function, and the switching penalty continuous term in the nonlinear optimization problem, and establish the iterative initialization structure for the sequential quadratic programming algorithm; In the iterative initialization structure, write the initial value vectors of the cold and heat source operating state variables, the initial value vectors of the equipment continuous operation variables, and the initial value vectors of the set parameters. The initial value vectors are arranged in time index order to form the control variable initial value sequence. In the current iteration, gradient calculation is performed on the objective function at the initial value sequence of control variables, and Jacobian matrix calculation is performed on the constraint set at the initial value sequence of control variables. The gradient calculation results and Jacobian matrix calculation results are written into the quadratic approximation buffer. In the current iteration, the switching penalty continuous term is calculated by differential chaining at the initial value sequence of the control variables to obtain the switching penalty gradient sequence. The switching penalty gradient sequence and the objective function gradient are written into the quadratic approximation buffer. A quadratic approximation objective function is constructed in the quadratic approximation buffer. The quadratic approximation objective function consists of the objective function gradient, the switching penalty gradient sequence, and the quadratic approximation Hessian matrix. Linearization constraints are constructed in the quadratic approximation buffer. The linearization constraints consist of the Jacobian matrix of the constraint condition set and the constraint residual vector. Solve the quadratic programming subproblem by applying the quadratic approximation objective function and linearization constraints, and output the incremental sequence of control variables. Perform a step scaling operation on the control variable increment sequence, and the step scaling operation outputs the step scaling coefficient to generate a scaled increment sequence; Update the initial value sequence of control variables using the scaling increment sequence, and write the update result into the initial value sequence of control variables for the next iteration. After each iteration, the change in the objective function value and the norm of the constraint residual vector are calculated. When the termination threshold is met, the final control variable update sequence is output and written into the control variable update sequence buffer.
7. The air conditioning temperature non-intrusive control method based on online learning according to claim 6, characterized in that, S6 includes the following steps: The control variable update vectors are read from the control variable update sequence buffer in chronological order. A vector summation operation is performed between the control variable update vectors and the initial control variable value vectors at the corresponding time indices to generate a control variable result vector. A variable type parsing operation is performed on the control variable result vector, splitting it into sub-vectors of cold / heat source operating status variables, equipment continuous operation variables, and set parameter sub-vectors. A state mapping operation is performed on the cold / heat source operating status variable sub-vectors to generate cold / heat source operating mode identifiers. An amplitude constraint operation is performed on the equipment continuous operation variable sub-vectors to form an equipment operating frequency sequence. A parameter parsing operation is performed on the set parameter sub-vectors to obtain set parameter values. The cold / heat source operating mode identifiers, equipment operating frequency sequences, and set parameter values are combined in chronological order to form the air conditioning operation control strategy, which is then written into the operation control strategy buffer.
8. The air conditioning temperature non-sensory control method based on online learning according to claim 7, characterized in that, Specifically, S7 includes: At the end of the control cycle, the control cycle termination time index is located from the multi-source parameter time series. The indoor environmental parameters and equipment operating parameters at the corresponding time index are read, and the actual cooling and heating load status values are obtained according to the parameter position parsing rules. The predicted cooling and heating load status values corresponding to the termination time index are read from the cooling and heating load status prediction sequence. The difference operation is performed between the predicted cooling and heating load status values and the actual cooling and heating load status values to generate the prediction deviation value. If the prediction sequence covers several time index positions, the predicted cooling and heating load status values before the termination time index position and the actual cooling and heating load status values at the corresponding time index positions are read. The difference operation is performed on each time index, and the values are arranged in chronological order to form the prediction deviation sequence. The amplitude mapping operation is performed on the prediction deviation sequence to generate the deviation measurement sequence. The deviation measurement sequence is written into the prediction deviation buffer as input data for the multi-source reliability gating factor update operation.