ACU intelligent control system

Abstract

The invention discloses an ACU intelligent control system, and relates to the technical field of adaptive control, and the system comprises a prediction module which is used for deploying a multi-parameter sensing network, collecting environment data in real time, transmitting the environment data to a data processing unit, combining meteorological data and historical operation data, building a load prediction model based on deep learning, and achieving the precise prediction of a system load. The control module constructs a multi-objective optimization model including energy consumption minimization, comfort maximization, equipment service life maximization and renewable energy utilization rate maximization based on the predicted load sequence, and calculates an optimal control strategy through an improved dynamic programming algorithm to realize balance control; and the optimization module issues the optimized control strategy to each execution unit through the central controller, and dynamically updates the prediction model parameters and the control strategy to form a closed-loop optimization control system. The control precision and the energy utilization efficiency of the ACU system can be effectively improved, the environment change is captured in time, and the response speed under the abnormal working condition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grids, and in particular to an ACU intelligent control method and system. Background Art

[0002] As the scale of key infrastructure such as data centers continues to expand, the energy consumption of the supporting air handling unit (ACU) system continues to increase, and traditional ACU control systems have many shortcomings in actual operation. Existing ACU control systems generally adopt PID control strategies based on fixed sampling cycles, which cannot effectively cope with conditions with rapid load changes, and the control parameters are relatively fixed, making it difficult to adapt to dynamic changes in equipment characteristics and environmental conditions. At the same time, traditional control strategies often focus solely on temperature and humidity control accuracy or system energy consumption, and lack comprehensive consideration of multiple goals such as comfort, equipment life, and renewable energy utilization, resulting in low overall system operation efficiency.

[0003] Most existing systems use simple threshold control or a single prediction model with limited prediction accuracy, and are unable to fully utilize multi-source information such as historical operating data and meteorological data for system optimization. In terms of load forecasting, traditional methods mainly rely on statistical models or simple machine learning algorithms, which are difficult to capture the complex nonlinear characteristics of load changes, affecting the foresight and accuracy of control strategies. At the system optimization level, existing solutions often use offline optimization or fixed-cycle updates, lack real-time adaptive optimization mechanisms, and are unable to respond to system state changes and external environmental disturbances in a timely manner, restricting the energy-saving potential and operational reliability of the ACU system. Therefore, it is urgent to develop an intelligent ACU control system to improve the system's control accuracy, energy efficiency, and operational reliability. Summary of the invention

[0004] In view of the problems existing in the existing ACU intelligent control method, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention lies in the technical problems of the existing ACU control system in terms of insufficient load prediction accuracy, difficulty in capturing load changes, and lack of real-time adaptive optimization mechanism.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides an ACU intelligent control system, which includes: a prediction module, which deploys a multi-parameter sensor network, collects environmental data in real time and transmits it to a data processing unit, combines meteorological data and historical operation data, establishes a load prediction model based on deep learning, and realizes accurate prediction of system load; a control module, which constructs a multi-objective optimization model including minimization of energy consumption, maximization of comfort, maximization of equipment life and maximization of renewable energy utilization based on the predicted load sequence, calculates the optimal control strategy through an improved dynamic programming algorithm, and realizes balanced control under multi-objective constraints; an optimization module, which is based on the optimized control strategy and is sent to each execution unit through a central controller. At the same time, the system continuously collects operation data for adaptive optimization, dynamically updates the prediction model parameters and control strategy, and forms a closed-loop optimization control system.

[0008] As a preferred solution of the ACU intelligent control system described in the present invention, wherein: the prediction module obtains real-time data of temperature, humidity and power through the deployed sensor network, and combines meteorological forecast data and historical operation data, uses deep learning methods to establish a load prediction model, and provides prediction information for the control module; the control module constructs a multi-objective optimization model based on the load sequence output by the prediction module, and uses an improved dynamic programming algorithm to find the optimal balance point among multiple goals of minimizing energy consumption, maximizing comfort, maximizing equipment life and maximizing renewable energy utilization, and generates a specific control strategy sequence; the optimization module converts the strategy generated by the control module into specific execution instructions and sends them to each execution unit, and feeds back the continuously collected operation data to the prediction module and the control module, so as to realize dynamic update of model parameters and control strategies.

[0009] In the second aspect, an embodiment of the present invention provides an ACU intelligent control method, which includes: deploying a multi-parameter sensor network, collecting environmental data in real time, combining meteorological data and historical operation data, establishing a load prediction model based on deep learning, and realizing accurate prediction of system load; constructing a multi-objective optimization model based on the predicted load sequence, including minimizing energy consumption, maximizing comfort, maximizing equipment life, and maximizing renewable energy utilization, and calculating the optimal control strategy through an improved dynamic programming algorithm to achieve balanced control under multi-objective constraints; based on the optimized control strategy, it is sent to each execution unit through the central controller, and the operation data is continuously collected for adaptive optimization, and the prediction model parameters and control strategy are dynamically updated to form a closed-loop optimization control.

[0010] As a preferred solution of the ACU intelligent control method described in the present invention, the data processing unit preprocesses the collected temperature parameters, humidity parameters, carbon dioxide concentration parameters and personnel density parameters, including data cleaning and standardization processing; the preprocessed environmental parameters are combined with historical operation data and meteorological characteristic data in the meteorological database to perform multi-scale feature extraction and data fusion; the data processing unit constructs a multi-dimensional feature vector system to form a basic feature vector with temperature parameters, humidity parameters, carbon dioxide concentration parameters and personnel density parameters, and encodes the time information of sampling time, date, week, month and season into a time feature vector to establish an associated expression between environmental parameters and time dimension; the statistical features of the data are calculated within the sliding time window, including the mean reflecting the trend of the data set, the standard deviation representing the degree of data dispersion, and the peak factor describing the data fluctuation characteristics, and the statistical quantities are composed of statistical feature vectors to represent the overall distribution characteristics of the data; the basic feature vectors, time feature vectors and statistical feature vectors are aligned and merged based on the timestamp to form a complete feature data set and provide a unified feature expression.

[0011] As a preferred solution of the ACU intelligent control method of the present invention, the prediction model is constructed based on the feature data set by adopting a hybrid structure of a convolutional neural network and a long short-term memory network to output high-precision prediction data; the prediction model extracts features in the feature data set through two one-dimensional convolutional layers, and the convolution operation is expressed as:

[0012] F 1 =ReLU(Conv1D(X,W 1 )+b 1 )

[0013] F 2 =ReLU(Conv1D(F 1 ,W 2 )+b 2 )

[0014] Among them, X is the input feature data set, W 1 ,W 2 is the convolution kernel weight matrix, b 1 ,b 2 are the corresponding bias vectors, F 1 ,F 2 They are the feature map outputs of the first and second layers respectively; the convolution output of each layer is batch normalized and expressed as:

[0015]

[0016] Among them, BN(x) represents the result of normalizing and rescaling the input data x, μ Bis the mean of the batch data, is the variance of the batch data, γ is a learnable scaling parameter, β is a learnable translation parameter, and ε is a numerical stability coefficient; based on F 2 The feature map calculates dual attention weights, including spatial attention weights and temporal attention weights; the calculation of the spatial attention weight is:

[0017] α(t)=softmax(V·tanh(W s ·F 2 (t)+b s ))

[0018] Among them, α(t) represents the spatial attention weight coefficient at time t, W s represents the spatial weight matrix, V represents the projection matrix, b s represents the bias vector; the calculation of the temporal attention weight is:

[0019]

[0020] c(t)=softmax(e(i,t))·F 2 (i)

[0021] Among them, e(i,t) represents the attention score of time i to the current time t, F 2 (i) represents the feature representation at time i, F 2 (t) represents the feature representation at the current time t, W T represents the time weight matrix, represents the transposed matrix of the time weight matrix, b T represents the bias vector of temporal attention, represents the transpose of the time projection vector, c(t) represents the time attention weight coefficient at time t; the weighted feature a is obtained by combining the weighted features by combining the spatial attention and the time attention. t , expressed as:

[0022]

[0023]

[0024] in, is the feature after spatial attention weighting, a t is the feature after attention weighting; the weighted feature is input into the two-layer bidirectional LSTM network, and the first layer is expressed as:

[0025] h t =BiLSTM1(a t ,h t-1 )

[0026] H 1 =[h 1 ,h 2 ,...,h t ]

[0027] Among them, a t is the attention weighted feature at the current time t, h t-1 is the hidden state of the previous moment, h t is the bidirectional LSTM output at the current moment, BiLSTM1 is the first layer of bidirectional LSTM network; H 1 is the output sequence of the first layer; the second layer is represented as:

[0028] o t =BiLSTM2(H 1 ,o t-1 )

[0029] Among them, t-1 is the hidden state of the second layer at the previous moment, o t The output of the second layer at the current moment, BiLSTM2 is the second-layer bidirectional LSTM network; integrating the bidirectional time series information, the final output is the predicted load sequence Y = {y 1 ,y 2 ,...,y n}.

[0030] As a preferred solution of the ACU intelligent control method of the present invention, the corresponding objective function is constructed based on the predicted load sequence for the four goals of minimizing energy consumption, maximizing comfort, maximizing equipment life and maximizing renewable energy utilization; the objective function Je of minimizing energy consumption is expressed as:

[0031] Je=α 1 ·Pc+α 2 ·Pf

[0032] Among them, α 1 and α 2 represents the weight coefficient, Pc represents the compressor cooling power, and Pf represents the fan delivery power; the objective function Jc for maximizing the comfort level is expressed as:

[0033] Jc=β 1 ·(TT s ) 2 +β 2 (HH s ) 2

[0034] Where, T is the actual room temperature, T s is the set temperature, H is the actual relative humidity, H s To set the humidity, β1 , β 2 is the temperature and humidity weight coefficient; the objective function Ji of maximizing the equipment life is expressed as:

[0035] Ji=γ 1 Ns+γ 2 ·Rt

[0036] Among them, Ns represents the number of starts and stops per unit time, Rt represents the cumulative running time, γ 1 and γ 2 is the life impact weight coefficient; the objective function Jr of maximizing the utilization rate of renewable energy is expressed as:

[0037] Jr=δ 1 ·(Pg-Pc-Pf) / Pg

[0038] Among them, δ 1 is the renewable energy utilization weight coefficient, Pg is the predicted power of photovoltaic power generation; the four sub-functions of minimizing energy consumption, maximizing comfort, maximizing equipment life and maximizing renewable energy utilization are linearly combined through the comprehensive objective function, which is expressed as:

[0039] J=w 1 ·Je+w 2 ·Jc+w 3 ·Ji+w 4 ·Jr

[0040] Among them, J represents the comprehensive objective function, w i is the weight coefficient, based on the relative deviation of load prediction δ p Dynamic adjustment is performed; the weight coefficient is adjusted in the form of sigmoid function, which is expressed as:

[0041]

[0042] Among them, k i is the weight adjustment coefficient.

[0043] As a preferred solution of the ACU intelligent control method of the present invention, the dynamic programming algorithm adopts the state space partition method to define the system state vector as s = [T, H, Pc, Pf], the temperature space [T min ,T max ] is divided into m parts, humidity space [H min ,H max ] is divided into n parts, forming an m×n state grid; in adjacent states s i to j In the process of state transfer, the state transfer cost g(s i ,s j) is calculated by weighting the energy consumption change ΔE, comfort change ΔC and life loss ΔL, and is expressed as:

[0044] g(s i ,s j )=μ 1 ·ΔE+μ 2 ·ΔC+μ 3 ΔL

[0045] ΔE=|E(s j )-E(s i )|

[0046] ΔC=|C(s j )-C(s i )|

[0047] ΔL=|L(s j )-L(s i )|

[0048] Among them, ΔE represents the change of energy consumption, E(s) is the total energy consumption of the system corresponding to state s, ΔC represents the change of comfort, C(s) is the PMV index corresponding to state s, ΔL represents the change of life loss, L(s) is the equipment loss corresponding to state s, μ 1 , μ 2 , μ 3 is the weight coefficient; in order to speed up the optimal path search, a heuristic evaluation function is constructed, which is expressed as:

[0049] h(s)=r 1 ·d(s,g)+r 2 ·e(s)

[0050]

[0051] e(s)=η·(Pc+Pf)·Δt

[0052] Among them, h(s) is the estimated cost from the current state s to the target state g, d(s,g) is the Euclidean distance from the current state s to the target state g, e(s) is the estimated energy consumption from the current state to the target state, r 1 ,r 2 is the balance coefficient, η is the energy consumption prediction coefficient; through the state transition cost g(s i ,s j ) and heuristic evaluation function h(s), construct the evaluation function of the A* algorithm to select the optimal state transfer path, expressed as:

[0053] f(s)=G(s)+h(s)

[0054] Among them, G(s) is the cumulative cost from the starting state to the current state s, and the cumulative state transfer cost g(s i ,s j ) is obtained, h(s) is the estimated cost from the current state to the target state; the search process uses the A* algorithm to search for the optimal path, and the output optimal state transfer path is used as the basis for generating specific control instructions.

[0055] As a preferred solution of the ACU intelligent control method of the present invention, wherein: the control sequence in the prediction time domain [t, t+N] is calculated based on the optimal state transfer path, including generating a cooling capacity adjustment instruction according to the temperature and humidity deviation, which is expressed as:

[0056] Qc(t)=f(T(t),H(t),T s ,H s )

[0057] Where T(t) is the actual indoor temperature at the current time t, H(t) is the actual relative humidity at the current time t, and T s Target setting temperature, H s The target set humidity, Qc(t) is the output cooling capacity control command; the fan speed command is calculated based on the required delivery power, expressed as:

[0058]

[0059] Among them, Pf(t) is the fan power at the current time t, k is the fan characteristic coefficient, and ω(t) is the output fan speed command; the valve opening command is determined in combination with the cooling capacity demand, expressed as:

[0060] v(t)=Qc(t) / Q max

[0061] Where Qc(t) is the cooling capacity command at the current time t, Q max is the maximum cooling capacity of the system, v(t) is the output valve relative opening instruction; the combination forms a complete control instruction sequence, which is expressed as:

[0062] U(t)=[Qc(t),ω(t),v(t)]

[0063] Among them, U(t) is the output control instruction sequence; U(t) is used as the input signal of the system actuator to drive the ACU system to run according to the optimized trajectory.

[0064] As a preferred solution of the ACU intelligent control method of the present invention, the adaptive optimization includes: the main controller calculates the deviation vector e(t)=[T(t)-T s ,H(t)-H s] and the total energy consumption of the system E(t) = Pc(t) + Pf(t). If the temperature deviation exceeds the preset temperature deviation threshold or the humidity deviation exceeds the preset humidity deviation threshold, the comfort weight w 2 Increase according to the preset step size. If the total energy consumption of the system E(t) exceeds the preset energy consumption threshold, the energy saving weight w 1 Increase according to the preset step size, and update the objective function J at the preset time interval through the weight adaptive mechanism; the main controller is based on the historical data of the preset time window [T(t), H(t), Pc(t), Pf(t)], if the temperature prediction error of the state prediction model exceeds the preset temperature prediction error threshold or the humidity prediction error exceeds the preset humidity prediction error threshold, then start the recursive least squares algorithm to update the model parameter η, if the estimation error of the heuristic evaluation function h(s) exceeds the preset evaluation error threshold, then adjust the balance coefficient r according to the preset balance coefficient adjustment step size 1 ,r 2 The main controller recalculates the optimal path Path based on the updated objective function J and the prediction model. If the Euclidean distance deviation between the new path and the original path is less than the preset path deviation threshold, the control instruction U(t) is fine-tuned within the preset range. If the deviation exceeds the preset path deviation threshold, the control instruction sequence U(t) is completely updated to form an adaptive optimization control loop based on performance indicators.

[0065] The beneficial effect of the present invention is that the present invention maps the operating state of the system to a node in the state space, and constructs a multi-objective optimization function that takes into account comfort and energy consumption. In the actual operation process, the system dynamically adjusts the sampling frequency through an adaptive sampling mechanism based on real-time data collected by temperature and humidity sensors and power sensors, and continuously updates the prediction model parameters using a recursive least squares algorithm. It can effectively improve the control accuracy and energy utilization efficiency of the ACU system, capture environmental changes in a timely manner, and improve the response speed under abnormal conditions. The prediction model can continuously adapt to changes in equipment characteristics and environmental disturbances, improve the robustness of the control strategy, realize real-time optimization and adaptive adjustment of the control strategy, and significantly reduce system energy consumption while improving user comfort. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0067] Figure 1 Flow chart of the ACU intelligent control method. DETAILED DESCRIPTION

[0068] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0069] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0070] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0071] Example 1

[0072] Reference Figure 1 , which is the first embodiment of the present invention, provides an ACU intelligent control method, including:

[0073] S1: By deploying a multi-parameter sensor network system, collecting environmental data in real time, combining meteorological data and historical operation data, and establishing a load prediction model based on deep learning, accurate prediction of system load can be achieved.

[0074] A multi-parameter sensing network system is constructed by deploying a variety of sensors, including temperature sensors, humidity sensors, carbon dioxide concentration sensors and personnel density sensors.

[0075] The temperature sensor uses PT100 platinum resistance, with a range of -20℃~50℃, an accuracy of ±0.1℃, and a sampling frequency of 1 time / minute;

[0076] The humidity sensor uses a capacitive sensor with a range of 0 to 100% RH, an accuracy of ±1% RH, and a sampling frequency of 1 time / 5 minutes;

[0077] The carbon dioxide concentration sensor uses an infrared sensor with a range of 0 to 5000ppm, an accuracy of ±50ppm, and a sampling frequency of 1 time / 5 minutes;

[0078] The personnel density sensor uses an infrared array sensor with a detection range of 100 square meters, an accuracy of ±1 person, and a sampling frequency of 1 time / 5 minutes.

[0079] The collected environmental parameters are collected in the data acquisition module and transmitted to the data processing unit to ensure the real-time and transmission stability of the data.

[0080] The collected temperature parameters, humidity parameters, carbon dioxide concentration parameters and personnel density parameters are preprocessed, including data cleaning and standardization.

[0081] The linear interpolation method is used to process the null values ​​in data cleaning, the five-point cubic smoothing filter is used to process the noise, and the outlier processing is based on the 3σ principle, where σ is the standard deviation.

[0082] The cleaned temperature parameters, humidity parameters, carbon dioxide concentration parameters and personnel density parameters are normalized to the maximum and minimum values.

[0083] The preprocessed environmental parameters are combined with historical operation data and meteorological characteristic data in the meteorological database to perform multi-scale feature extraction and data fusion.

[0084] The data processing unit sets up multi-scale time windows. Specifically, the complete daily cycle variation characteristics are captured through the 24-hour main time window, and the short-term fluctuation characteristics are captured through the 1-hour sub-time window. A 15-minute sliding step is used to achieve continuous data collection to ensure the temporal integrity and real-time performance of feature extraction.

[0085] By constructing a multi-dimensional feature vector system, temperature parameters, humidity parameters, carbon dioxide concentration parameters and personnel density parameters are combined into basic feature vectors. At the same time, the sampling time, date, week, month, season and other time information are encoded into time feature vectors to establish the association expression between environmental parameters and time dimension.

[0086] The data processing unit calculates the statistical characteristics of the data within the sliding time window, including the mean value reflecting the trend of the data set, the standard deviation representing the degree of data dispersion, and the peak factor describing the data fluctuation characteristics, and combines the statistics into a statistical feature vector to represent the overall distribution characteristics of the data;

[0087] Based on the timestamp, the basic feature vectors, time feature vectors and statistical feature vectors are aligned and merged to form a complete feature data set, realizing the time series fusion of multi-source heterogeneous data and providing a unified feature expression.

[0088] Based on the complete feature data set, a prediction model is constructed using a hybrid structure of convolutional neural network and long short-term memory network to output high-precision prediction data.

[0089] The prediction model extracts features from the feature dataset through two one-dimensional convolutional layers. The convolution operation is expressed as:

[0090] F 1=ReLU(Conv1D(X,W 1 )+b 1 )

[0091] F 2 =ReLU(Conv1D(F 1 ,W 2 )+b 2 )

[0092] Among them, X is the input feature data set, W 1 ,W 2 is the convolution kernel weight matrix, b 1 ,b 2 are the corresponding bias vectors, F 1 ,F 2 They are the feature map outputs of the first and second layers respectively.

[0093] The first layer uses 64 convolution kernels to extract basic feature patterns, and the second layer uses 128 convolution kernels to extract combined features. The output of each convolution layer is batch normalized and expressed as:

[0094]

[0095] Among them, BN(x) represents the result of normalizing and rescaling the input data x, μ B is the mean of the batch data, is the variance of the batch data, γ is a learnable scaling parameter, β is a learnable translation parameter, and ε is a numerical stability coefficient.

[0096] Based on F 2 The feature map calculates dual attention weights, including spatial attention weights and temporal attention weights. The calculation of spatial attention weights is:

[0097] α(t)=softmax(V·tanh(W s ·F 2 (t)+b s ))

[0098] Among them, α(t) represents the spatial attention weight coefficient at time t, W s represents the spatial weight matrix, V represents the projection matrix, b s Represents the bias vector.

[0099] The temporal attention weight is calculated as:

[0100]

[0101] c(t)=softmax(e(i,t))·F 2 (i)

[0102] Among them, e(i,t) represents the attention score of time i to the current time t, F 2 (i) represents the feature representation at time i, F 2 (t) represents the feature representation at the current time t, W T represents the time weight matrix, used for feature transformation, represents the transposed matrix of the time weight matrix, b T represents the bias vector of temporal attention, represents the transpose of the time projection vector, and c(t) represents the time attention weight coefficient at time t.

[0103] The attention scores are converted into probability weights through softmax, and the relevant historical information is adaptively selected by weighting the importance of historical features to generate contextual representations containing temporal dependencies.

[0104] Furthermore, the weighted features are combined by combining spatial attention and temporal attention to obtain the weighted feature a t , expressed as:

[0105]

[0106] in, is the feature after spatial attention weighting, a t It is the attention-weighted feature, which contains information in both spatial and temporal dimensions.

[0107] The weighted features are input into a two-layer bidirectional LSTM network, and the first layer is expressed as:

[0108] h t =BiLSTM1(a t ,h t-1 )

[0109] H 1 =[h 1 ,h 2 ,...,h t ]

[0110] Among them, a t is the attention weighted feature at the current time t, combining the information of spatial and temporal attention, h t-1 is the hidden state of the previous moment, h t is the bidirectional LSTM output at the current moment, including the forward and backward hidden state concatenation. BiLSTM1 is the first layer of the bidirectional LSTM network, containing 128 memory units. 1 is the output sequence of the first layer, collecting the hidden state outputs at all times.

[0111] The second layer is represented as:

[0112] o t =BiLSTM2(H 1 ,o t-1 )

[0113] Among them, t-1 is the hidden state of the second layer at the previous moment, o t The output of the second layer at the current moment. BiLSTM2 is the second-layer bidirectional LSTM network, which contains 64 memory units.

[0114] The first layer captures long-term dependencies through 128 memory units, and the second layer uses 64 memory units to integrate bidirectional timing information using the same calculation method, and finally outputs the predicted load sequence Y = {y 1 ,y 2 ,...,y n}.

[0115] The prediction model is optimized and trained using a composite loss function of three combinations, where the composite loss function is designed using a weighted combination of MSE, MAE, and Huber loss, expressed as:

[0116] Loss = λ 1 MSE+λ 2 MAE+λ 3 Huber

[0117] Among them, λ 1 , 2 , 3 are the weight coefficients of each loss item, MSE is the mean square error, which is used to measure the overall prediction error, MAE is the mean absolute error, which improves the robustness of the model to outliers, and Huber loss achieves a smooth transition between MSE and MAE. The threshold parameter of Huber loss is used to control the sensitivity of the loss function.

[0118] The optimizer uses the Adam algorithm, the initial learning rate is set to 0.001, and the momentum parameter β is set 1 =0.9,β 2 = 0.999 for adaptive adjustment of gradient updates;

[0119] During the training process, the batch size is set to 64 to balance the computational efficiency and the accuracy of gradient estimation. The maximum number of training rounds is set to 100 to avoid overtraining. An early stopping strategy is adopted to terminate the training when the validation loss has not improved for 5 consecutive rounds. At the same time, the model parameters with the lowest validation loss are saved. Finally, the standardized prediction sequence and its reliability evaluation report are output to provide reliable data support for load balancing decisions.

[0120] S2: Based on the predicted load sequence, a multi-objective optimization model is constructed, which includes minimizing energy consumption, maximizing comfort, maximizing equipment life, and maximizing renewable energy utilization. The optimal control strategy is calculated through an improved dynamic programming algorithm to achieve balanced control under multi-objective constraints.

[0121] Based on the predicted load sequence Y={y 1 ,y 2 ,...,y n}Construct corresponding objective functions for the four core goals of minimizing energy consumption, maximizing comfort, maximizing equipment life and maximizing renewable energy utilization.

[0122] The energy consumption objective function Je is expressed by a weighted combination of the compressor cooling power Pc and the fan delivery power Pf:

[0123] Je=α 1 ·Pc+α 2 ·Pf

[0124] Among them, α 1 and α 2 represents the weight coefficient, Pc represents the compressor cooling power, Pf represents the fan delivery power, and the weight coefficient α 1 and α 2 Used to balance the energy consumption contribution of cooling and air supply.

[0125] Furthermore, the compressor refrigeration power Pc is obtained by multiplying the refrigeration coefficient COP and the refrigeration capacity Qc, which is expressed as:

[0126] Pc=COP·Qc

[0127] The fan delivery power Pf is determined based on the fan characteristic coefficient k and the speed ω, and is expressed as:

[0128] Pf=k·ω 3

[0129] The comfort objective function Jc is calculated by comparing the actual temperature T with the set temperature T s The squared deviation of the actual humidity H and the set humidity H s The squared deviation term is composed of:

[0130] Jc=β 1 ·(TT s ) 2 +β 2 (HH s ) 2

[0131] Where, T is the actual room temperature, T s is the set temperature, H is the actual relative humidity, H sTo set the humidity, β 1 , β 2 is the temperature and humidity weight coefficient. Weight coefficient β 1 and β 2 Used to adjust the priority of temperature and humidity control.

[0132] The equipment life objective function Ji combines the number of equipment starts and stops per unit time Ns and the cumulative operating time Rt, expressed as:

[0133] Ji=γ 1 Ns+γ 2 ·Rt

[0134] Among them, Ns represents the number of starts and stops per unit time, Rt represents the cumulative running time, γ 1 and γ 2 is the lifespan influence weight coefficient. 1 and γ 2 Balance the impact of start-stop losses and operating losses.

[0135] The renewable energy objective function Jr is expressed as:

[0136] Jr=δ 1 ·(Pg-Pc-Pf) / Pg

[0137] Among them, δ 1 is the weight coefficient of renewable energy utilization, and Pg is the predicted power of photovoltaic power generation.

[0138] Based on the requirements for system operation safety and equipment protection, multiple operation constraints are set, including: the range of indoor temperature T is determined by the upper and lower limits T min and T max Limit to ensure basic comfort requirements; the control range of indoor relative humidity H is determined by H min and H max Constraints to avoid air being too dry or humid; minimum compressor running time t on Not less than T minon , minimum downtime t off Not less than T minoff , to prevent damage to the equipment caused by frequent starts and stops; the number of starts and stops per unit time Ns is limited to the maximum allowable value Ns max , protect the compressor; the system power change rate |ΔP| is limited to ΔP max To ensure the smooth operation of the system and reduce the peak power impact.

[0139] Based on the load forecast deviation, the adaptive weight adjustment is performed, and the four sub-functions of minimizing energy consumption, maximizing comfort, maximizing equipment life, and maximizing renewable energy utilization are linearly combined through the comprehensive objective function, which is expressed as:

[0140] J=w 1 ·Je+w 2 ·Jc+w 3 ·Ji+w 4 ·Jr

[0141] Among them, w i is the weight coefficient, based on the relative deviation of load prediction δ p Make dynamic adjustments.

[0142] The adjustment of the weight coefficient adopts the form of sigmoid function, which is expressed as:

[0143]

[0144] Among them, k i is the weight adjustment coefficient, which is determined through experimental optimization.

[0145] If the prediction accuracy is high, the weight of energy saving and renewable energy utilization is increased through the weight function; if the prediction deviation is large, the weight of comfort control is increased to achieve closed-loop optimization of prediction-control; the weight coefficient is normalized to ensure the numerical stability of the optimization process;

[0146] The improved dynamic programming algorithm of state space partition is adopted to define the system state vector as s = [T, H, Pc, Pf], the temperature space [T min ,T max ] is divided into m parts, humidity space [H min ,H max ] is divided into n parts to form m×n state grids, which transforms the continuous control problem into a discrete decision problem and reduces the computational complexity.

[0147] In the adjacent state s i to j In the process of state transfer, the state transfer cost g(s i ,s j ) is calculated by weighting the energy consumption change ΔE, comfort change ΔC and life loss ΔL, and is expressed as:

[0148] g(s i ,s j )=μ 1 ·ΔE+μ 2 ·ΔC+μ 3 ΔL

[0149] ΔE=|E(s j )-E(s i )|

[0150] ΔC=|C(s j )-C(s i)|

[0151] ΔL=|L(s j )-L(s i )|

[0152] Among them, ΔE represents the change of energy consumption, E(s) is the total energy consumption of the system corresponding to state s, ΔC represents the change of comfort, C(s) is the PMV index corresponding to state s, ΔL represents the change of life loss, L(s) is the equipment loss corresponding to state s, μ 1 , μ 2 , μ 3 is the weight coefficient.

[0153] In order to speed up the optimal path search, a heuristic evaluation function is constructed, which is expressed as:

[0154] h(s)=r 1 ·d(s,g)+r 2 ·e(s)

[0155]

[0156] e(s)=η·(Pc+Pf)·Δt

[0157] Among them, h(s) is the estimated cost from the current state s to the target state g, d(s,g) is the Euclidean distance from the current state s to the target state g, e(s) is the estimated energy consumption from the current state to the target state, r 1 ,r 2 is the balance coefficient, which is used to weigh the importance of path length and energy consumption estimation. η is the energy consumption prediction coefficient, which is obtained through historical data statistics.

[0158] Through the state transfer cost g(s i ,s j ) and heuristic evaluation function h(s), construct the evaluation function of the A* algorithm to select the optimal state transfer path, effectively reducing the computational complexity, expressed as:

[0159] f(s)=G(s)+h(s)

[0160] Among them, G(s) is the cumulative cost from the starting state to the current state s, and the cumulative state transfer cost g(s i ,s j ) is obtained, h(s) is the estimated cost from the current state to the target state.

[0161] The search process uses the A* algorithm to search for the optimal path. Specifically:

[0162] The initial state s 0Add to the open list (open table), and leave the closed table empty. In each iteration, select the state s with the smallest evaluation function f(s) from the open table as the current expansion node, remove it from the open table and add it to the closed table. Then expand all feasible adjacent states s' of the current state s:

[0163] Calculate the actual cumulative cost G(s') from the starting state to s' via the current state s, and the estimated cost h(s') from s' to the target state, so as to obtain the total evaluation value f(s') of s'.

[0164] For each adjacent state s', if s' is not in the open table and the closed table, add it to the open table; if s' is already in the open table, compare the cost of the new and old paths and retain the path with the lower cost; if s' is already in the closed table and the new path has a lower cost, add s' back to the open table.

[0165] Repeat the above expansion process until the current state selected from the open table reaches the target state, or the open table is empty (indicating no feasible path). The optimal state transfer path finally output will serve as the basis for generating specific control instructions.

[0166] The optimal state transfer path is obtained to calculate and predict the control sequence in the time domain [t, t+N], including generating cooling capacity adjustment instructions according to temperature and humidity deviations, which can be expressed as:

[0167] Qc(t)=f(T(t),H(t),T s ,H s )

[0168] Where T(t) is the actual indoor temperature at the current time t, H(t) is the actual relative humidity at the current time t, and T s Target setting temperature, H s The target set humidity, Qc(t) is the output cooling capacity control instruction.

[0169] The fan speed command is calculated based on the required delivery power and is expressed as:

[0170]

[0171] Wherein, Pf(t) is the fan power at the current time t, k is the fan characteristic coefficient, and ω(t) is the output fan speed command.

[0172] The valve opening instruction is determined in combination with the cooling capacity demand, expressed as:

[0173] v(t)=Qc(t) / Q max

[0174] Where Qc(t) is the cooling capacity command at the current time t, Q max is the maximum cooling capacity of the system, and v(t) is the output relative valve opening instruction.

[0175] These basic control quantities are combined to form a complete control instruction sequence, which is expressed as:

[0176] U(t)=[Qc(t),ω(t),v(t)]

[0177] Among them, U(t) is the output control instruction sequence.

[0178] U(t) is used as the input signal of the system actuator to drive the ACU system to run according to the optimized trajectory.

[0179] S3: The optimized control strategy is sent to each execution unit through the central controller, and the operation data is continuously collected for adaptive optimization, and the prediction model parameters and control strategy are dynamically updated to form a closed-loop optimization control.

[0180] The main controller outputs a control instruction sequence U(t) according to the optimal state transfer path Path generated by the objective function J. If a system startup signal is detected, the main controller increases the slope of the cooling capacity instruction Qc(t) to the calculated value; if the system is in a normal operating state, the control instruction sequence U(t) is sent to the variable frequency compressor, the electronic expansion valve and the variable frequency fan through the field bus respectively;

[0181] The temperature and humidity sensors and power sensors in the system collect operating parameters [T(t), H(t), Pc(t), Pf(t)] according to the preset sampling period. If the temperature T(t) change rate or the humidity H(t) change rate exceeds the preset temperature change threshold or the preset humidity change threshold, the fast sampling mode is triggered to reduce the sampling period to the preset minimum value. If the parameter fluctuation falls back to the normal range, the normal sampling period is restored, and the collected data is transmitted back to the main controller through the field bus;

[0182] The main controller calculates the deviation vector e(t)=[T(t)-T s ,H(t)-H s ] and the total energy consumption of the system E(t) = Pc(t) + Pf(t). If the temperature deviation exceeds the preset temperature deviation threshold or the humidity deviation exceeds the preset humidity deviation threshold, the comfort weight w 2 Increase according to the preset step size. If the total energy consumption of the system E(t) exceeds the preset energy consumption threshold, the energy saving weight w 1 Increase according to the preset step size, and update the objective function J at the preset time interval through the weight adaptive mechanism;

[0183] Based on the historical data of the preset time window [T(t), H(t), Pc(t), Pf(t)], the main controller starts the recursive least squares algorithm to update the model parameter η if the temperature prediction error of the state prediction model exceeds the preset temperature prediction error threshold or the humidity prediction error exceeds the preset humidity prediction error threshold. If the estimation error of the heuristic evaluation function h(s) exceeds the preset evaluation error threshold, the balance coefficient r is adjusted according to the preset balance coefficient adjustment step size. 1 ,r 2 ;

[0184] The main controller recalculates the optimal path Path based on the updated objective function J and the prediction model. If the Euclidean distance deviation between the new path and the original path is less than the preset path deviation threshold, the control instruction U(t) is fine-tuned within the preset range. If the deviation exceeds the preset path deviation threshold, the control instruction sequence U(t) is completely updated to form an adaptive optimization control loop based on performance indicators.

[0185] In summary, the ACU intelligent control system provided by the present invention maps the operating state of the system to nodes in the state space by establishing a state space model and a heuristic evaluation function, and constructs a multi-objective optimization function that takes into account comfort and energy consumption. In the actual operation process, the system dynamically adjusts the sampling frequency through an adaptive sampling mechanism based on real-time data collected by temperature and humidity sensors and power sensors, and continuously updates the prediction model parameters using a recursive least squares algorithm. It can effectively improve the control accuracy and energy utilization efficiency of the ACU system, capture environmental changes in a timely manner, and improve the response speed under abnormal conditions. The prediction model can continuously adapt to changes in equipment characteristics and environmental disturbances, improve the robustness of the control strategy, realize real-time optimization and adaptive adjustment of the control strategy, and significantly reduce system energy consumption while improving user comfort.

[0186] This embodiment further provides an ACU intelligent control system, including:

[0187] The prediction module deploys a multi-parameter sensor network to collect environmental data in real time and transmit it to the data processing unit. It combines meteorological data and historical operation data to establish a load prediction model based on deep learning to achieve accurate prediction of system load.

[0188] The control module builds a multi-objective optimization model based on the predicted load sequence, including minimizing energy consumption, maximizing comfort, maximizing equipment life, and maximizing renewable energy utilization. It calculates the optimal control strategy through an improved dynamic programming algorithm to achieve balanced control under multi-objective constraints.

[0189] The optimization module sends the optimized control strategy to each execution unit through the central controller. At the same time, the system continuously collects operation data for adaptive optimization, dynamically updates the prediction model parameters and control strategy, and forms a closed-loop optimization control system.

[0190] This embodiment also provides a computer device suitable for the ACU intelligent control method, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the ACU intelligent control method proposed in the above embodiment.

[0191] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0192] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the ACU intelligent control method proposed in the above embodiment is implemented.

[0193] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0194] Example 2

[0195] This embodiment provides an ACU intelligent control method. In order to verify the beneficial effects of the present invention, a scientific demonstration is carried out through simulation experiments.

[0196] The ACU system of a data center was selected as the research object. The experimental period was 30 days, and the actual load data was used for simulation. The ambient temperature setting range was 20-26℃, and the relative humidity setting range was 40%-60%. The system configuration included a variable frequency compressor (cooling capacity 5-25kW adjustable), an electronic expansion valve, and a variable frequency fan (air volume 20%-100% adjustable).

[0197] Comparison plans include:

[0198] Solution 1: Traditional PID control (fixed sampling period 60s);

[0199] Solution 2: Single prediction model + fixed weight optimization control;

[0200] Solution 3: The intelligent control system of the present invention.

[0201] The experimental results are shown in Table 1:

[0202] Table 1 Comparison results of performance indicators of different control schemes

[0203]

[0204]

[0205] Furthermore, the response characteristics of typical load mutation conditions are analyzed and compared, as shown in Table 2:

[0206] Table 2 Comparison of response characteristics of typical load mutation conditions

[0207] Response characteristics Traditional PID control Single prediction + fixed weight Solution of the present invention Sampling period(s) 60 30 10 (Adaptive) Temperature fluctuation range (℃) ±1.5 ±0.8 ±0.5 Adjustment time (min) 8.5 6.8 5.7 Steady-state COP 3.5 3.8 4.2

[0208] From the comparative data in Table 1, it can be seen that the scheme of the present invention has a comprehensive performance improvement compared with the traditional PID control and the single prediction fixed weight scheme. In terms of temperature and humidity control accuracy, the mean temperature control deviation is reduced to ±0.5℃, which is 43.2% better than the traditional scheme, and the mean humidity control deviation is reduced to ±2.8%, which is 38.7% higher, indicating that the control accuracy is significantly improved; in terms of energy efficiency, the average daily energy consumption of the system is reduced to 381kWh, which is 21.5% lower than the traditional scheme, and the average COP of the system is increased to 4.2, an increase of 18.3%; in terms of dynamic performance, the adjustment time under load mutation is shortened by 52.6%, and the temperature overshoot is reduced by 47.8%, reflecting a better dynamic response characteristic; in terms of prediction performance, the 24-hour load prediction accuracy is increased to 92.3%, which is 31.5% higher than the traditional scheme.

[0209] Table 2 further shows the response characteristics of the system under typical load mutation conditions. The scheme of the present invention dynamically adjusts the sampling period to 10s through an adaptive sampling mechanism, so that the temperature fluctuation is controlled within the range of ±0.5℃, while maintaining a high COP value of 4.2, which fully demonstrates that the present invention can still achieve the unity of precise control and efficient operation under extreme conditions. These data strongly prove the significant advantages of the present invention in improving the control performance and operating efficiency of the ACU system.

[0210] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. ACU intelligent control system, characterized by: include: The prediction module deploys a multi-parameter sensor network to collect environmental data in real time and transmit it to the data processing unit. It combines meteorological data and historical operation data to establish a load prediction model based on deep learning to achieve accurate prediction of system load. The control module builds a multi-objective optimization model based on the predicted load sequence, including minimizing energy consumption, maximizing comfort, maximizing equipment life, and maximizing renewable energy utilization. It calculates the optimal control strategy through an improved dynamic programming algorithm to achieve balanced control under multi-objective constraints. The optimization module sends the optimized control strategy to each execution unit through the central controller. At the same time, the system continuously collects operation data for adaptive optimization, dynamically updates the prediction model parameters and control strategy, and forms a closed-loop optimization control system.

2. The ACU intelligent control system according to claim 1, characterized in that: The prediction module obtains real-time data of temperature, humidity and power through the deployed sensor network, and uses deep learning methods to establish a load prediction model in combination with weather forecast data and historical operation data to provide prediction information for the control module; The control module constructs a multi-objective optimization model based on the load sequence output by the prediction module, and uses an improved dynamic programming algorithm to find the optimal balance point between multiple objectives of minimizing energy consumption, maximizing comfort, maximizing equipment life, and maximizing renewable energy utilization, and generates a specific control strategy sequence; The optimization module converts the strategy generated by the control module into specific execution instructions and sends them to each execution unit, and feeds back the continuously collected operation data to the prediction module and the control module to realize the dynamic update of model parameters and control strategies.

3. An ACU intelligent control method, based on the ACU intelligent control system according to any one of claims 1 to 2, characterized in that: include: Deploy a multi-parameter sensor network to collect environmental data in real time, combine meteorological data and historical operation data, and establish a load prediction model based on deep learning to achieve accurate prediction of system load; Based on the predicted load sequence, a multi-objective optimization model including minimizing energy consumption, maximizing comfort, maximizing equipment life and maximizing renewable energy utilization is constructed. The optimal control strategy is calculated through an improved dynamic programming algorithm to achieve balanced control under multi-objective constraints. The optimized control strategy is sent to each execution unit through the central controller. At the same time, the operation data is continuously collected for adaptive optimization, and the prediction model parameters and control strategy are dynamically updated to form a closed-loop optimization control.

4. The ACU intelligent control method according to claim 3, characterized in that: The data processing unit pre-processes the collected temperature parameters, humidity parameters, carbon dioxide concentration parameters and personnel density parameters, including data cleaning and standardization processing; The pre-processed environmental parameters are combined with historical operation data and meteorological characteristic data in the meteorological database to perform multi-scale feature extraction and data fusion; The data processing unit constructs a multi-dimensional feature vector system, combines temperature parameters, humidity parameters, carbon dioxide concentration parameters and personnel density parameters into basic feature vectors, and encodes the time information of sampling time, date, week, month and season into a time feature vector, so as to establish an associated expression between environmental parameters and time dimension; Calculate the statistical characteristics of the data within the sliding time window, including the mean that reflects the trend of the data set, the standard deviation that represents the degree of data dispersion, and the peak factor that describes the data fluctuation characteristics. The statistics are combined into a statistical feature vector to represent the overall distribution characteristics of the data. Based on the timestamp, the basic feature vector, time feature vector and statistical feature vector are aligned and merged to form a complete feature data set and provide a unified feature expression.

5. The ACU intelligent control method according to claim 4, characterized in that: Based on the feature data set, a prediction model is constructed using a hybrid structure of a convolutional neural network and a long short-term memory network to output high-precision prediction data; The prediction model extracts features from the feature data set through two one-dimensional convolutional layers. The convolution operation is expressed as: F1 = ReLU (Conv1D (X, W1) + b1) F2 = ReLU (Conv1D (F1, W2) + b2) Among them, X is the input feature data set, W1, W2 are the convolution kernel weight matrices, b1, b2 are the corresponding bias vectors, F1, F2 are the feature map outputs of the first and second layers respectively; The output of each convolution layer is batch normalized and expressed as: Among them, BN(x) represents the result of normalizing and rescaling the input data x, μ B is the mean of the batch data, is the variance of the batch data, γ is a learnable scaling parameter, β is a learnable translation parameter, and ε is a numerical stability coefficient; Calculate dual attention weights based on the F2 feature map, including spatial attention weights and temporal attention weights; The spatial attention weight is calculated as: α(t)=softmax(V·tanh(W s ·F2(t)+b s )) Among them, α(t) represents the spatial attention weight coefficient at time t, W s represents the spatial weight matrix, V represents the projection matrix, b s represents the bias vector; The temporal attention weight is calculated as: c(t)=softmax(e(i,t))·F2(i) Among them, e(i,t) represents the attention score of time i to the current time t, F2(i) represents the feature representation of time i, F2(t) represents the feature representation of the current time t, and W T represents the time weight matrix, represents the transposed matrix of the time weight matrix, b T represents the bias vector of temporal attention, represents the transpose of the time projection vector, c(t) represents the time attention weight coefficient at time t; Combining spatial attention and temporal attention, we can obtain weighted features by combining weighted features. t , expressed as: in, is the feature after spatial attention weighting, a t The feature after attention weighting; The weighted features are input into a two-layer bidirectional LSTM network, and the first layer is expressed as: h t =BiLSTM1(a t ,h t-1 ) H1=[h1,h2,...,h t ] Among them, a t is the attention weighted feature at the current time t, h t-1 is the hidden state of the previous moment, h t is the bidirectional LSTM output at the current moment, BiLSTM1 is the first-layer bidirectional LSTM network; H1 is the first-layer output sequence; The second layer is represented as: o t =BiLSTM2(H1,o t-1 ) Among them, t-1 is the hidden state of the second layer at the previous moment, o t The output of the second layer at the current moment, BiLSTM2 is the second-layer bidirectional LSTM network; integrating the bidirectional time series information, the final output is the predicted load sequence Y = {y1, y2, ..., y n }.

6. The ACU intelligent control method according to claim 5, characterized in that: Based on the predicted load sequence, a corresponding objective function is constructed for the four goals of minimizing energy consumption, maximizing comfort, maximizing equipment life, and maximizing renewable energy utilization; The objective function Je of minimizing energy consumption is expressed as: Je=α1·Pc+α2·Pf Among them, α1 and α2 represent weight coefficients, Pc represents the compressor cooling power, and Pf represents the fan delivery power; The objective function Jc of maximizing the comfort level is expressed as: Jc=β1·(T-T s ) 2 +β2(H-H s ) 2 Where T is the actual room temperature, T s is the set temperature, H is the actual relative humidity, H s is the set humidity, β1 and β2 are the temperature and humidity weight coefficients; The objective function Ji of maximizing the equipment life is expressed as: Ji=γ1·Ns+γ2·Rt Among them, Ns represents the number of starts and stops per unit time, Rt represents the cumulative running time, and γ1 and γ2 are the life impact weight coefficients; The objective function Jr for maximizing the utilization rate of renewable energy is expressed as: Jr=δ1·(Pg-Pc-Pf) / Pg Among them, δ1 is the weight coefficient of renewable energy utilization, and Pg is the predicted power of photovoltaic power generation; The four sub-functions of minimizing energy consumption, maximizing comfort, maximizing equipment life, and maximizing renewable energy utilization are linearly combined through the comprehensive objective function, which is expressed as: J=w1·Je+w2·Jc+w3·Ji+w4·Jr Among them, J represents the comprehensive objective function, w i is the weight coefficient, based on the relative deviation of load prediction δ p Make dynamic adjustments; The adjustment of the weight coefficient adopts the form of sigmoid function, which is expressed as: Among them, k i is the weight adjustment coefficient.

7. The ACU intelligent control method according to claim 6, characterized in that: The dynamic programming algorithm adopts the state space partition method to define the system state vector as s = [T, H, Pc, Pf], the temperature space [T min ,T max ] is divided into m parts, humidity space [H min ,H max ] is divided into n parts, forming m×n state grids; In the adjacent state s i to j In the process of state transfer, the state transfer cost g(s i ,s j ) is calculated by weighting the energy consumption change ΔE, comfort change ΔC and life loss ΔL, and is expressed as: g(s i ,s j )=μ1·ΔE+μ2·ΔC+μ3·ΔL ΔE=|E(s j )-E(s i )| ΔC=|C(s j )-C(s i )| ΔL=|L(s j )-L(s i )| Among them, ΔE represents the change in energy consumption, E(s) is the total system energy consumption corresponding to state s, ΔC represents the change in comfort, C(s) is the PMV index corresponding to state s, ΔL represents the change in life loss, L(s) is the equipment loss corresponding to state s, μ1, μ2, μ3 are weight coefficients; In order to speed up the optimal path search, a heuristic evaluation function is constructed, which is expressed as: h(s)=r1·d(s,g)+r2·e(s) e(s)=η·(Pc+Pf)·Δt Among them, h(s) is the estimated cost from the current state s to the target state g, d(s,g) is the Euclidean distance from the current state s to the target state g, e(s) is the estimated energy consumption from the current state to the target state, r1, r2 are balance coefficients, and η is the energy consumption prediction coefficient; Through the state transfer cost g(s i ,s j ) and heuristic evaluation function h(s), construct the evaluation function of the A* algorithm to select the optimal state transfer path, expressed as: f(s)=G(s)+h(s) Among them, G(s) is the cumulative cost from the starting state to the current state s, and the cumulative state transfer cost g(s i ,s j ) is obtained, h(s) is the estimated cost from the current state to the target state; The search process uses the A* algorithm to search for the optimal path, and the output optimal state transfer path is used as the basis for generating specific control instructions.

8. The ACU intelligent control method according to claim 7, characterized in that: The control sequence in the prediction time domain [t, t+N] is calculated based on the optimal state transfer path, including generating a cooling capacity adjustment instruction according to the temperature and humidity deviation, which is expressed as: Qc(t)=f(T(t),H(t),T s ,H s ) Where T(t) is the actual indoor temperature at the current time t, H(t) is the actual relative humidity at the current time t, and T s Target setting temperature, H s Target set humidity, Qc(t) is the output cooling capacity control instruction; The fan speed command is calculated based on the required delivery power and is expressed as: Wherein, Pf(t) is the fan power at the current time t, k is the fan characteristic coefficient, and ω(t) is the output fan speed command; The valve opening instruction is determined in combination with the cooling capacity demand, expressed as: v(t)=Qc(t) / Q max Where Qc(t) is the cooling capacity command at the current time t, Q max is the maximum cooling capacity of the system, v(t) is the output valve relative opening instruction; Combined to form a complete control instruction sequence, expressed as: U(t)=[Qc(t),ω(t),v(t)] Among them, U(t) is the output control instruction sequence; U(t) is used as the input signal of the system actuator to drive the ACU system to run according to the optimized trajectory.

9. The ACU intelligent control method according to claim 8, characterized in that: The adaptive optimization includes: The main controller calculates the deviation vector e(t)=[T(t)-T s ,H(t)-H s ] and the total energy consumption of the system E(t) = Pc(t) + Pf(t). If the temperature deviation exceeds the preset temperature deviation threshold or the humidity deviation exceeds the preset humidity deviation threshold, the comfort weight w2 is increased by a preset step size. If the total energy consumption of the system E(t) exceeds the preset energy consumption threshold, the energy-saving weight w1 is increased by a preset step size. The objective function J is updated at a preset time interval through the weight adaptive mechanism. Based on the historical data of the preset time window [T(t), H(t), Pc(t), Pf(t)], the main controller starts the recursive least squares algorithm to update the model parameter η if the temperature prediction error of the state prediction model exceeds the preset temperature prediction error threshold or the humidity prediction error exceeds the preset humidity prediction error threshold. If the estimation error of the heuristic evaluation function h(s) exceeds the preset evaluation error threshold, the balance coefficients r1 and r2 are adjusted according to the preset balance coefficient adjustment step size. The main controller recalculates the optimal path Path based on the updated objective function J and the prediction model. If the Euclidean distance deviation between the new path and the original path is less than the preset path deviation threshold, the control instruction U(t) is fine-tuned within the preset range. If the deviation exceeds the preset path deviation threshold, the control instruction sequence U(t) is completely updated to form an adaptive optimization control loop based on performance indicators.

Images