Integrated controller intelligent analysis and management method and system

Abstract

The application discloses an integrated controller intelligent analysis and management method and system, relates to the field of integrated control, and utilizes historical data to complete complex air conditioning system identification model training and offline optimization of an optimal PID parameter strategy in a cloud with abundant computing power, generates an expert strategy library containing various working conditions and corresponding optimal control parameters. Subsequently, a lightweight proxy model with small calculation amount and fast reasoning speed is trained based on the strategy library. The proxy model is deployed on a resource-limited edge controller. In this way, the edge end does not need to perform complex online optimization calculation, and only needs to input real-time working condition data into the proxy model, so that the current optimal PID parameter can be quickly reasoned. The architecture of cloud training and edge reasoning realizes online, real-time and intelligent self-setting of the controller parameter, and effectively bridges the gap between high-level algorithms and industrial landing.

Description

Technical Field

[0001] This application relates to the field of integrated control, and more specifically, to an intelligent analysis and management method and system for integrated controllers. Background Technology

[0002] With the increasing scale and complexity of modern buildings and industrial facilities, integrated controllers, as the core of automation systems, directly impact overall energy consumption and operational stability through their management and control efficiency. Taking large central air conditioning systems as an example, their operating conditions are dynamically influenced by various factors such as indoor and outdoor temperature and humidity, pedestrian traffic, and solar radiation intensity, exhibiting highly nonlinear and time-varying characteristics. Traditional controllers often employ fixed PID (proportional-integral-derivative) parameters, which are ill-suited to complex and changing actual operating conditions, leading to system lag or overshoot, resulting in unnecessary energy waste and equipment damage. Therefore, to achieve refined energy efficiency management and adaptive control, constructing a solution capable of intelligently analyzing and managing integrated controllers, enabling them to self-tune optimal control parameters based on real-time operating conditions, has become an urgent need for industry development.

[0003] In existing technological explorations, researchers have attempted to utilize artificial intelligence algorithms to optimize controller parameters. A common approach is to upload massive amounts of operating data to a cloud server, leveraging the cloud's powerful computing capabilities to run complex optimization models (such as deep neural networks or reinforcement learning models) to calculate optimal PID parameters, and then sending the results down to the field controller for execution. However, this cloud-based decision-making and terminal-based execution model presents a core contradiction: the pursuit of highly intelligent decision-making through complex algorithms clashes significantly with the inherent characteristics of industrial field controllers—low computing power, low latency, and high reliability. On one hand, control loops have extremely high real-time requirements, and the communication latency between data and the cloud is unacceptable for many rapid-response scenarios. On the other hand, industrial controllers (such as DDCs and PLCs), due to cost and reliability considerations, typically employ resource-constrained embedded processors, which cannot directly support and run computationally intensive AI models. This gap between algorithms and deployment makes it difficult to effectively implement advanced intelligent strategies in industrial settings, resulting in significant engineering bottlenecks for the intelligent management of integrated controllers in practice.

[0004] Therefore, an innovative cloud-edge collaborative architecture is desired. Summary of the Invention

[0005] To resolve the conflict between real-time constraints and computational resources for complex models, according to one aspect of this application, an integrated controller intelligent analysis and management method is provided, comprising: In the cloud, acquire historical sensor data streams from the air conditioning system; In the cloud, the historical sensor data stream of the air conditioning system is subjected to operating condition feature engineering to obtain an air conditioning operating condition feature engineering dataset. In the cloud, a system identification model is trained based on the air conditioning operating condition feature engineering dataset to obtain a trained system identification model. In the cloud, the optimal PID controller parameters are generated based on the trained system identification model to obtain the optimal policy dataset; In the cloud, a lightweight supervised learning model is trained on the optimal policy dataset to obtain an agent policy model, which is then deployed at the edge. At the edge, acquire real-time sensor data; At the edge, the real-time sensor data is input into the agent strategy model to obtain updated PID parameters.

[0006] According to another aspect of this application, an integrated controller intelligent analysis and management system is provided, comprising: The historical sensor data stream acquisition module is used to acquire historical sensor data streams of the air conditioning system in the cloud. The operating condition feature engineering module is used to perform operating condition feature engineering on the historical sensor data stream of the air conditioning system in the cloud to obtain an air conditioning operating condition feature engineering dataset. The system identification model training module is used to train the system identification model in the cloud based on the air conditioning operating condition feature engineering dataset to obtain the trained system identification model. The optimal strategy dataset generation module is used to generate optimal PID controller parameters in the cloud based on the trained system identification model to obtain the optimal strategy dataset. The agent strategy model training module is used to train a lightweight supervised learning model on the cloud based on the optimal policy dataset to obtain an agent strategy model, which is deployed at the edge. The real-time sensor data acquisition module is used to acquire real-time sensor data at the edge. The PID parameter update module is used to input the real-time sensor data into the agent strategy model at the edge to obtain the updated PID parameters.

[0007] Compared with existing technologies, this application provides an integrated controller intelligent analysis and management method and system, which separates the deployment of computationally intensive tasks from tasks with high real-time requirements. Specifically, in the cloud with abundant computing power, historical data is used to complete the training of a complex air conditioning system identification model and the offline optimization of the optimal PID parameter strategy, generating an expert strategy library containing various operating conditions and their corresponding optimal control parameters. Subsequently, a lightweight proxy model with low computational cost and fast inference speed is trained based on this strategy library. This proxy model is deployed on a resource-constrained edge controller. In this way, the edge does not need to perform complex online optimization calculations; it only needs to input real-time operating condition data into the proxy model to quickly infer the current optimal PID parameters. This cloud-based training and edge-based inference architecture cleverly solves the core contradiction mentioned in the background technology between complex AI algorithms and the low computing power and high real-time requirements of field controllers, realizing online, real-time, and intelligent self-tuning of controller parameters, effectively bridging the gap between advanced algorithms and industrial applications. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a flowchart of an integrated controller intelligent analysis and management method according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of data flow in the integrated controller intelligent analysis and management method according to an embodiment of this application.

[0011] Figure 3 This is a flowchart of step 2 in the integrated controller intelligent analysis and management method according to an embodiment of this application.

[0012] Figure 4 This is a flowchart of step 4 in the integrated controller intelligent analysis and management method according to an embodiment of this application.

[0013] Figure 5 This is a block diagram of an integrated controller intelligent analysis and management system according to an embodiment of this application. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0015] In response to the problems mentioned in the background description, this application proposes an intelligent analysis and management method for integrated controllers. Figure 1 This is a flowchart of an integrated controller intelligent analysis and management method according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the integrated controller intelligent analysis and management method according to an embodiment of this application. Figure 1 and Figure 2 As shown, the integrated controller intelligent analysis and management method according to an embodiment of this application includes: Step 1, acquiring historical sensor data streams of an air conditioning system in the cloud; Step 2, performing operating condition feature engineering on the historical sensor data streams of the air conditioning system in the cloud to obtain an air conditioning operating condition feature engineering dataset; Step 3, training a system identification model based on the air conditioning operating condition feature engineering dataset in the cloud to obtain a trained system identification model; Step 4, generating optimal PID controller parameters based on the trained system identification model in the cloud to obtain an optimal strategy dataset; Step 5, training a lightweight supervised learning model based on the optimal strategy dataset in the cloud to obtain a proxy strategy model, wherein the proxy strategy model is deployed at the edge; Step 6, acquiring real-time sensor data at the edge; Step 7, inputting the real-time sensor data into the proxy strategy model at the edge to obtain updated PID parameters.

[0016] In step 1, historical sensor data streams of the air conditioning system are acquired in the cloud. It should be understood that to achieve intelligent optimization of the integrated controller's control strategy, the model first needs to deeply understand the dynamic characteristics of the controlled object (i.e., the air conditioning system). The operating behavior of the air conditioning system is influenced by a variety of internal and external factors, exhibiting complex nonlinearity and time-varying characteristics, whose inherent laws cannot be accurately described by simple physical formulas. Historical sensor data streams are a true record of the system's past operating states, containing the inherent correlations between system inputs, outputs, and states under different operating conditions. Therefore, acquiring and analyzing massive amounts of historical sensor data streams in the cloud lays the data foundation for building a system identification model that can accurately simulate the dynamic response of a real air conditioning system. This model is the prerequisite and key for subsequent optimal control strategy simulation and optimization.

[0017] In one possible implementation, step 1 is performed as follows: Specifically, historical data is stored in a time-series database, such as an instance based on a surge database or a time-stamped database. This database stores multiple data tables from the central air conditioning system of a large shopping mall over the past year, recording values ​​from different sensors and actuators at different points in time. When performing the data acquisition operation, a pre-defined data query request is sent to this database. This request specifies the time range to be acquired, for example, from 00:00:00 on January 1, 2022 to 23:59:59 on December 31, 2022, and specifies the key data fields to be extracted.

[0018] These fields constitute the historical raw sensor data stream, and its content may include, but is not limited to: basic data used to calculate operating condition characteristics, such as environmental and indoor state quantities like outdoor temperature and indoor average temperature, as well as the set temperature as the control target. Additionally, to enable subsequent construction of a system identification model and evaluation of control effectiveness, the data stream also includes key equipment status and controlled variables, especially equipment power consumption data. Each data point is accompanied by a precise timestamp.

[0019] For example, a query request might be structured as a Structured Query Language (SQL) statement that filters data from a database for a specified time range and fields. After the query is executed, the database returns one or more datasets. These datasets are integrated and processed to form a unified, time-indexed raw data stream. This data stream can be a collection of multiple time series, each corresponding to a sensor or control signal. For example, one time series might be the indoor average temperature, consisting of a series of (timestamp, temperature value) data pairs. Finally, the acquired historical sensor data stream is loaded into the memory or temporary storage of a cloud server for use in subsequent condition characterization steps.

[0020] In step 2, on the cloud, the historical sensor data stream of the air conditioning system undergoes operating condition feature engineering to obtain an air conditioning operating condition feature engineering dataset. Correspondingly, while the raw historical sensor data stream obtained from the database contains rich information, its format is fragmented and unprocessed, making it unusable by machine learning models. This raw data often suffers from noise, missing values, and misaligned timestamps, and its numerical values ​​may not directly reveal the deeper physical meaning of the air conditioning's operating state, such as the system's thermal inertia and load change trends. To enable subsequent system identification models to more accurately learn and predict the dynamic behavior of the air conditioning, the raw data needs to be deeply processed and refined. Therefore, to transform the raw, unstructured time-series data into structured feature vectors that accurately and comprehensively describe the air conditioning's operating conditions at any given time, thereby constructing a high-quality air conditioning operating condition feature engineering dataset suitable for model training, this application performs operating condition feature engineering.

[0021] In one possible implementation, Figure 3 This is a flowchart of step 2 in the integrated controller intelligent analysis and management method according to an embodiment of this application. Figure 3 As shown, step 2 involves performing condition feature engineering on the historical sensor data stream of the air conditioning system in the cloud to obtain an air conditioning condition feature engineering dataset. This includes: step 21, preprocessing the historical sensor data stream of the air conditioning system to obtain a cleaned and aligned time series table; step 22, constructing and vectorizing the condition features of the cleaned and aligned time series table to obtain multiple feature columns, and merging the multiple feature columns with the cleaned and aligned time series table to obtain an enhanced feature table; step 23, using the feature row vector at time point t in the enhanced feature table as the i-th training sample; step 24, using a key controlled variable at time point t+1 in the enhanced feature table as the label of the i-th training sample; and step 25, merging the i-th training sample and its label to obtain a training sample in the air conditioning condition feature engineering dataset.

[0022] In the above implementation scheme, step 2 is performed as follows: The first stage is step 21, data preprocessing. First, data cleaning is performed. The input data stream is loaded into the memory of the cloud computing environment, and each time series, such as the indoor average temperature series, is iterated and checked. The cleaning process includes outlier handling and missing value imputation. For outliers, a reasonable physical threshold range can be preset; for example, the effective range of indoor temperature is set to 10 degrees Celsius to 40 degrees Celsius. When a data point's value (such as -99 or 200) is detected to be outside this range, the data point is identified as an outlier and replaced with an invalid label, such as NaN. For missing values, including null values ​​in the original data and outliers marked as invalid, linear interpolation is used for imputation. For example, if the temperature at time point t-1 is 25 degrees Celsius, the temperature at time point t+1 is 26 degrees Celsius, and the temperature at time point t is missing, then the temperature at time t is calculated using linear interpolation to be 25.5 degrees Celsius, and this value is used for imputation. Next, time alignment is performed. Because different sensors (such as thermometers and power meters) may have different sampling frequencies and reporting cycles, the timestamps of each data stream are not entirely consistent. To construct a unified analytical view, a standard time sampling interval needs to be set; in this embodiment, this interval is set to 5 minutes. Then, a resampling operation is performed on all cleaned time series. Specifically, data points within each 5-minute window are aggregated. For example, for an indoor average temperature series with an original sampling frequency of 1 minute, the average of the 5 data points within each 5-minute window is calculated as the representative value for that window; similarly, for a device power consumption series with an original sampling frequency of 10 seconds, the average of all power consumption readings within each 5-minute window is calculated. All series are processed in this way to ensure that each series has a unique corresponding value at the same time point, such as 10:00, 10:05, 10:10... In particular, the above preset values, such as the effective range of indoor temperature and the time sampling interval, are determined based on a comprehensive consideration of the physical characteristics of the air conditioning system, industry standards and specifications, and historical data distribution statistical analysis. After completing the above operations, all the cleaned and time-aligned sequences are combined into a single table. The index of this table is a unified 5-minute interval timestamp, and each column corresponds to a sensor or control signal, such as outdoor temperature, indoor average temperature, device power consumption, etc., which is the cleaned and aligned time sequence table.

[0023] The second stage, step 22, involves the construction and vectorization of operating condition characteristics. In one possible implementation, the characteristics in the multiple feature columns include temperature error, building heat load, time lag characteristics of equipment power consumption, average equipment power consumption, and equipment power consumption change rate. It is worth noting that while the raw data is authentic, its information is isolated and superficial, failing to directly reveal the inherent logic of the control process, the inertia of the physical system, and the combined impact of external environmental disturbances. An effective model needs to understand these deep-seated, dynamically changing operating condition characteristics. Therefore, a series of derived characteristics with clear physical meaning and dynamic representation capabilities need to be calculated. Temperature error quantifies the deviation of the current control effect from the target and is a core indicator for evaluating control performance. Building heat load explicitly expresses the impact of the external environment (such as outdoor temperature) on internal energy consumption demand. Introducing the time lag characteristics of equipment power consumption and average equipment power consumption captures and characterizes the inertial effect of air conditioning operation and the recent energy consumption baseline level. Calculating the equipment power consumption change rate is to characterize the instantaneous trend of load changes. By combining these characteristics, the operating conditions of the air conditioning system at any given moment can be comprehensively and three-dimensionally depicted from multiple dimensions such as control deviation, external disturbance, dynamic inertia, energy consumption benchmark, and changing trends.

[0024] Specifically, firstly, based on the existing columns in the cleaned and aligned time series table, new feature columns are calculated and generated one by one. The specific calculation process is as follows: traverse each row of the time series table (i.e., each time point), read the indoor average temperature value and the set temperature value of that row, subtract the two, and use the result as the new feature temperature error value for that row. For example, if the indoor average temperature of a row is 25.5 degrees Celsius and the set temperature is 24 degrees Celsius, then the temperature error value for that row is 1.5. Similarly, traverse each row and read the outdoor temperature and indoor average temperature values. The heat load is estimated using a preset simplified heat transfer formula: Building heat load = C * (Outdoor temperature - Indoor average temperature). Here, the heat transfer coefficient C is a constant predetermined based on the specific physical parameters of the building envelope, area, etc., for example, it can be set to 1.2. If the outdoor temperature of a row is 32 degrees Celsius and the indoor average temperature is 25.5 degrees Celsius, then the building heat load value for that row is 1.2 * (32 - 25.5) = 7.8. To introduce dynamic information from the time series, the equipment power consumption column is shifted. For example, to generate the "Device Power Consumption - Lag 1 Step" feature, shift the entire "Device Power Consumption" column down one row, so that the feature value for row t is the power consumption value of row t-1 in the original table. Similarly, generating "Device Power Consumption - Lag 2 Step" and "Device Power Consumption - Lag 3 Step" features corresponds to shifting the original power consumption column down two and three rows, respectively. The "Device Average Power Consumption" feature is obtained by calculating a moving average. A time window is set, for example, 30 minutes (i.e., 6 sampling points). For row t, its "Device Average Power Consumption" value is the arithmetic mean of the 6 values ​​from row t-5 to row t in the original "Device Power Consumption" column. The "Device Power Consumption Change Rate" feature represents the instantaneous change in power consumption. For row t, its "Device Power Consumption Change Rate" value is the device power consumption value of that row minus the device power consumption value of row t-1. After calculating all the above new features, multiple newly generated data columns will be obtained. Subsequently, a column merging operation is performed, adding these new feature columns (temperature error, building heat load, time lag characteristics of equipment power consumption, average equipment power consumption, and rate of change of equipment power consumption) to the right side of the original cleaned and aligned time series table. The merged table is the enhanced feature table.

[0025] The third stage is step 23, constructing training samples. In one possible implementation, the key controlled variable is device power consumption. Specifically, a traversal program is initiated, scanning the enhanced feature table row by row in chronological order. This table uses time as the row index and various original and derived features as columns, completely recording the operating conditions at all time points. When the program processes time point t, i.e., row t of the table, it reads the values ​​of all feature columns in that row and combines these values ​​into an ordered numerical vector. The dimension of this vector is the same as the number of columns in the enhanced feature table, and the order of its elements is consistent with the order of the columns. For example, at time point t, the program reads the following values ​​for this row: [Outdoor temperature: 32.0, Indoor average temperature: 25.5, Set temperature: 24.0, Device power consumption: 50.2, Temperature error: 1.5, Building heat load: 7.8, Device power consumption lag by 1 step: 48.9,...]. This complete row vector containing all feature values ​​is defined as the i-th training sample. Specifically, the program moves to the next row in the table, i.e., time point t+1, and repeats the same extraction process to generate the (i+1)th training sample. This process continues until all valid time rows in the augmented feature table have been traversed. The final result is a set of multiple feature row vectors, where each vector is an independent training sample waiting to be paired with a label.

[0026] The fourth stage, step 24, involves constructing training labels. In one possible implementation, the key controlled variable is device power consumption. That is, in the intelligent management of air conditioning systems, the core optimization objective typically focuses on maximizing energy efficiency while ensuring comfort. Device power consumption, as the most direct and crucial quantitative indicator of total system energy consumption, directly reflects the immediate impact of control strategy adjustments on energy consumption. Therefore, selecting device power consumption as the key controlled variable allows the subsequent system identification model to accurately learn and predict energy consumption responses under different operating conditions and control inputs, thereby providing a direct and effective predictive basis for generating energy-saving-oriented optimal control strategies.

[0027] Specifically, when processing the i-th training sample corresponding to time point t, the program queries the data at time point t+1 in the augmented feature table. Then, from the row at t+1, the program precisely locates the column named "Device Power Consumption" and reads the value of that cell. This value represents the actual power consumption of the air conditioner in the next time step. For example, if the i-th training sample is generated based on the feature row vector at time point 10:00, the program will look for the row at time point 10:05 in the augmented feature table. If the value of the "Device Power Consumption" column in that row is 51.3 kWh, then the value 51.3 is extracted and defined as the label corresponding to the i-th training sample. This process is synchronized with the sample generation in the previous step. Whenever a sample at time point t is generated, it is immediately matched with a label at time point t+1.

[0028] Finally, step 25 generates the final dataset. The program creates a new data structure, such as a data pair or a row in a data frame. This data structure contains two parts: the first part is the training sample, and the second part is its corresponding label. For example, continuing the previous example, the i-th training sample is the vector [32.0,25.5,24.0,50.2,1.5,7.8,48.9,...], and its corresponding label is the value 51.3. The merging operation will construct them into a data pair ([32.0,25.5,24.0,50.2,1.5,7.8,48.9,...],51.3). This data pair is defined as a complete training sample in the air conditioning operating condition feature engineering dataset. This merging process is performed on all generated sample-label pairs. After each merging is completed, a record is added to the final dataset. Once all the data from all time points has been processed, a collection containing a large number of such data pairs is formed, namely the air conditioning operating condition characteristic engineering dataset. Each record clearly defines the power consumption of the equipment at the next moment under what operating conditions.

[0029] In step 3, a system identification model is trained in the cloud based on the air conditioning operating condition characteristic engineering dataset to obtain the trained system identification model. It is understood that a core prerequisite for achieving intelligent optimization of controller parameters is the need for a digital twin capable of accurately simulating the dynamic behavior of a real air conditioner. This virtual environment can predict how key air conditioning indicators (such as device power consumption) will change under specific operating conditions and control inputs. Since the operating mechanism of air conditioning is extremely complex and difficult to describe with precise physical equations, a data-driven approach is needed to construct such a model. Therefore, this application utilizes a dataset to train a system identification model to construct a mathematical model capable of accurately reproducing and predicting the dynamic response of the air conditioner. This model will serve as a virtual experimental platform for subsequent control strategy simulation and optimization iteration.

[0030] In one possible implementation, step 3 is performed as follows: In this embodiment, the system identification model uses a Long Short-Term Memory (LSTM) network as its core architecture. A LSTM network is a special type of recurrent neural network whose internal structure includes input gates, forget gates, and output gates. It can effectively learn and remember long-term dependencies in time-series data, making it very suitable for simulating the dynamic processes of air conditioning with thermal inertia and time delay effects.

[0031] First, the model architecture is defined. The specific architecture of the constructed Long Short-Term Memory (LSTM) network model is as follows: Input layer, where the input dimension matches the feature vector dimension of the training samples. Before inputting the data into the model, normalization is performed, for example, using min-max scaling to scale all feature values ​​to the [0,1] interval to accelerate model convergence and improve stability. Long Short-Term Memory (LSTM) layer, which is the core of the model. In this embodiment, two LSTM layers can be stacked to enhance the model's learning ability. The first LSTM layer receives the normalized input feature vector, and its number of hidden units can be set to 128. The output of this layer is a sequence, which serves as the input to the second LSTM layer. The number of hidden units in the second LSTM layer can be set to 64. Each LSTM unit contains a series of weight matrices and bias vectors (e.g., parameters related to the input gate, forget gate, output gate, and cell state updates), which are randomly initialized at the start of training. After the output of the last LSTM layer, one or more fully connected layers are connected. In this embodiment, a fully connected layer containing 32 neurons is connected, and the Rectified Luminous Interval (ReLU) function is used as the activation function to nonlinearly combine the features extracted by the LSTM layer. Finally, a fully connected layer containing only a single neuron is connected as the output layer. This layer does not use an activation function (or uses a linear activation function), and its output is a direct prediction of the device's power consumption at the next time step.

[0032] Next, the model training process is carried out. First, the input air conditioning operating condition feature dataset is divided into a training set, a validation set, and a test set, for example, in an 8:1:1 ratio. The training set is used to calculate gradients and update the model's weight parameters; the validation set is used to monitor model performance during training, prevent overfitting, and adjust hyperparameters accordingly; the test set is used to finally evaluate the model's generalization ability after training is complete. The training process adopts a supervised learning approach. A loss function is set to measure the difference between the model's predicted values ​​and the true label values. For regression prediction tasks, mean squared error (MSE) is chosen as the loss function. At the same time, an optimizer algorithm is selected to update the model's weights and biases based on the gradients calculated by the loss function. For example, the Adaptive Moments Estimation (Adam) optimizer is used, and an initial learning rate is set, such as 0.001. Training is carried out iteratively, i.e., in batches and periods. In one period, the model traverses the entire training set once. In each batch, a small batch of data is taken from the training set (for example, the batch size is set to 64), and it is input into the model for forward propagation to obtain the predicted power consumption value. Next, the mean squared error between the predicted value and the actual power consumption label is calculated. Then, the gradient of the loss function with respect to all trainable parameters (weights and biases of each layer) in the model is calculated using the backpropagation algorithm. Finally, the Adam optimizer updates the parameters based on these gradients, shifting the value of the loss function in a decreasing direction. After each epoch, the current model is evaluated using a validation set. The loss value on the validation set is monitored; if the validation loss no longer decreases for several consecutive epochs, an early stopping mechanism can be triggered to terminate training and prevent overfitting. The entire training process continues until the preset maximum number of epochs is reached or early stopping is triggered.

[0033] After training, the model's internal weights and bias parameters are adjusted to an optimal or near-optimal state, enabling the model to make power consumption predictions that are closest to reality given historical operating conditions. This fixed model structure, containing the optimized parameters, is the trained system identification model. This model will be saved as a file for use in subsequent simulations and optimal strategy generation steps.

[0034] In step 4, on the cloud, optimal PID controller parameters are generated based on the trained system identification model to obtain an optimal strategy dataset. It should be understood that although a system identification model capable of accurately simulating air conditioning dynamics is available, this only solves the prediction problem, not the control problem—that is, how to find the optimal control strategy to match the ever-changing real-time operating conditions. Therefore, to build a lightweight model that can directly guide edge controllers in decision-making, an expert knowledge base is needed. This knowledge base should contain various typical operating conditions and their corresponding optimal control parameters. Therefore, by using the trained system identification model to build a virtual simulation environment and performing extensive offline simulations and optimizations through intelligent optimization algorithms, operating condition-optimal PID parameter data pairs covering various operating conditions can be generated systematically and efficiently, thereby constructing a high-quality optimal strategy dataset.

[0035] In one possible implementation, Figure 4 This is a flowchart of step 4 in the integrated controller intelligent analysis and management method according to an embodiment of this application. Figure 4 As shown, step 4, in the cloud, generates optimal PID controller parameters based on the trained system identification model to obtain the optimal strategy dataset, including: step 41, extracting the current operating condition feature vector from the air conditioning operating condition feature engineering dataset; step 42, constructing a configured simulation environment based on the trained system identification model and virtual PID controller; step 43, initializing a population containing multiple candidate PID parameter groups; step 44, injecting the candidate PID parameter groups into the configured simulation environment to obtain the control error time series and the device power consumption time series; step 45, inputting the control error time series and the device power consumption time series into a performance index function to obtain the fitness score of the candidate PID parameter group; step 46, selecting the candidate PID parameter group corresponding to the highest fitness score as the converged optimal PID parameter of the current operating condition feature vector; step 47, constructing a data pair between the current operating condition feature vector and the converged optimal PID parameter to obtain an optimal strategy data in the optimal strategy dataset.

[0036] In the above implementation scheme, step 4 is performed as follows: The first stage is step 41, which extracts representative operating conditions. In one possible implementation scheme, step 41, extracting the current operating condition feature vector from the air conditioning operating condition feature engineering dataset, includes: step 411, performing K-Means clustering analysis on the air conditioning operating condition feature engineering dataset to obtain K sampled operating condition vectors; step 412, extracting any one of the K sampled operating condition vectors as the current operating condition feature vector.

[0037] Specifically, the first step is to perform step 411, K-Means clustering analysis. K-Means clustering analysis is a classic unsupervised learning algorithm whose goal is to divide a dataset into K distinct, non-overlapping clusters, such that data points within each cluster are as similar as possible (close in distance), while data points between different clusters are as different as possible (far in distance). First, the number of clusters K needs to be determined. The value of K represents how many typical categories we want to divide all historical work conditions into. This value can be set based on prior knowledge, for example, by estimating a rough number of work condition categories based on combinations of factors such as season, weekday / holiday, day / night, etc. Alternatively, it can be determined through data-driven methods, such as the elbow rule or the silhouette coefficient method. In this embodiment, the clustering effect under different K values ​​is evaluated using the silhouette coefficient method, and a value that results in a high average silhouette coefficient is finally selected, for example, setting K=100. After determining the value of K, the K-means algorithm begins execution. The process is as follows: 1. Randomly select 100 feature vectors from all feature vectors in the air conditioning operating condition feature engineering dataset as initial cluster centers (centroids). 2. Traverse each remaining feature vector in the dataset and calculate its Euclidean distance to all 100 cluster centers. Assign the feature vector to the cluster represented by the nearest cluster center. 3. After all feature vectors have been assigned to clusters, recalculate the centroid of each cluster. The new centroid is the mean vector of all feature vectors within that cluster. Repeat steps 2 and 3 until the cluster assignments no longer change, at which point the algorithm converges. After convergence, the algorithm outputs two results: one is the cluster label of each original feature vector, and the other is the 100 finally determined cluster center vectors. Each of these 100 cluster center vectors is an average representation of all operating condition features within its cluster, and is therefore considered a high-level generalization of a class of similar operating conditions. The set of these 100 cluster center vectors is the K sampled operating condition vectors. The second step is to execute step 412 to extract the current operating condition feature vector from the sampled operating condition vectors. In the previous stage, 100 representative sampled load case vectors were obtained. During the actual optimal parameter generation, a loop or parallel task will be initiated to process these 100 sampled load case vectors one by one. In each loop, any unprocessed sampled load case vector is extracted from the set of these 100 vectors. This extracted vector is defined as the current load case feature vector for this optimization task. For example, in the first iteration of the loop, the first sampled load case vector C1 is extracted as the current load case feature vector. This vector C1 will serve as input to initiate the subsequent simulation environment construction and genetic algorithm optimization process. After optimizing C1, the loop enters the second iteration, extracting the second sampled load case vector C2 as the new current load case feature vector, and the optimization is restarted. This process will continue until all 100 sampled load case vectors have been processed as current load case feature vectors at least once.

[0038] The second stage, step 42, involves building the simulation environment. First, a virtual PID controller needs to be defined and implemented. This is a software-based algorithm module with functionality identical to the physical PID controller. This module receives an input value (control error) and calculates and outputs a control signal based on its three internally set parameters: proportional (Kp), integral (Ki), and derivative (Kd). Its core algorithm is a discrete-time PID control law: Output(t) = Kp*e(t) + Ki*Σe(t) + Kd*(e(t) - e(t-1)), where e(t) is the control error at the current moment, and e(t) = set temperature - indoor average temperature(t). The Kp, Ki, and Kd parameters of this virtual PID controller are variable; during the subsequent optimization process, the genetic algorithm will continuously inject different candidate parameter sets into it. Next, the simulation environment is built and configured. The construction process involves connecting the trained system identification model with the virtual PID controller to form a closed loop of information flow. The closed-loop operation logic is as follows: First, the simulation environment uses the current operating condition feature vector as its initial state. This vector contains all the necessary information at the start of the simulation (t=0), such as the initial indoor and outdoor temperatures, setpoint temperature, historical power consumption, etc. Next, the simulation progresses in discrete time steps (e.g., 5 minutes, consistent with the data sampling interval). At each simulation step t: 1. Extract the indoor average temperature and setpoint temperature from the current operating condition feature vector and calculate the control error e(t). 2. Input the control error e(t) into the virtual PID controller. The controller calculates a control signal output based on its currently configured Kp, Ki, Kd parameters; for example, this output could represent the target load rate of the chiller unit or the frequency of the chilled water pump. 3. Combine the control signal output by the virtual PID controller with other relevant features in the operating condition vector at the current time t, such as the outdoor temperature, to form a complete input vector. This input vector is provided to the trained system identification model. The model performs one forward propagation to predict the equipment power consumption and indoor average temperature at the next time t+1. 4. Update the operating condition vector using the model-predicted new equipment power consumption and average indoor temperature to form the operating condition vector for the next time step t+1. Simultaneously, update other time-dependent features, such as time lag features. 5. Record the control error at the current step size t and the predicted equipment power consumption. This stepping process continues, simulating a preset future duration, such as 2 hours, or 24 simulation steps. Finally, the output is a configured simulation environment. This environment is an executable software entity containing the following elements: an instance of a pre-trained system identification model; a virtual PID controller with dynamically configurable parameters that implements the PID algorithm; the current operating condition feature vector passed from step 41 configured as the starting point for the simulation; and a closed-loop simulation process that defines how the model and controller interact and how the state is updated.

[0039] The third stage is step 43, initialization. First, two key hyperparameters need to be set: population size and the value range of each PID parameter. The population size parameter determines how many candidate solutions are included in each generation. A larger population can better cover the solution space, but increases the computational cost of each generation. A smaller population is faster to compute, but may lead to insufficient diversity. In this embodiment, based on the trade-off between computational resources and optimization efficiency, the population size is set to 50. This means that each generation will have 50 different sets of PID parameters participating in competition and evolution. The parameter range needs to set a reasonable search boundary for each of the proportional (Kp), integral (Ki), and derivative (Kd) parameters. This range can be set based on prior knowledge of control engineering or preliminary analysis of historical operating data. Setting a range that is too wide will increase the search difficulty, while a range that is too narrow may exclude the optimal solution. In this embodiment, the parameter ranges are set as follows: Kp∈[0.1,20.0], Ki∈[0.01,2.0], Kd∈[0.0,10.0]. After determining the hyperparameters, initialization begins. The program executes a loop, the number of loops being the population size (50 times). In each loop, a candidate PID parameter set, also known as an individual or chromosome, is generated. The specific generation process is as follows: Within the preset range [0.1, 20.0], a random floating-point number is generated using a uniformly distributed random number generator as the Kp value for this individual. Within the preset range [0.01, 2.0], another random floating-point number is generated using a uniformly distributed random number generator as the Ki value for this individual. Within the preset range [0.0, 10.0], yet another random floating-point number is generated using a uniformly distributed random number generator as the Kd value for this individual. These three randomly generated parameter values ​​are combined into a triplet (Kp, Ki, Kd), which constitutes a complete candidate PID parameter set. For example, in the first iteration of the loop, the parameter set (12.5, 0.8, 3.2) might be generated; the second iteration might generate (2.1, 0.15, 7.8); and so on. The program will repeat this process 50 times, generating 50 distinct candidate PID parameter sets randomly distributed within a preset parameter space. Finally, these 50 generated parameter sets are aggregated into a set or list to obtain an initial population containing multiple candidate PID parameter sets.

[0040] The fourth stage is step 44, parameter injection. This process is executed through a loop, iterating through each candidate PID parameter set in the population. For each parameter set, for example, Kp=12.5, Ki=0.8, Kd=3.2, the following operations are performed: First, parameter injection is performed. This candidate PID parameter set (12.5, 0.8, 3.2) is set into the virtual PID controller module in the configured simulation environment. That is, in the subsequent simulation process, the virtual PID controller will use this specific set of Kp, Ki, Kd values ​​to calculate its control output. At the same time, the state of the simulation environment is reset to the initial current operating condition feature vector, ensuring that the starting point of each simulation is consistent. Second, the simulation is started and run. The simulation process is performed according to the preset step size and total duration. In this embodiment, the total simulation duration is set to 2 hours, the time step is 5 minutes, and there are a total of 24 simulation steps. In the first step (t=0 to t=5 minutes), the simulation environment starts from the initial operating condition. The program extracts the indoor average temperature and the set temperature from the initial operating condition feature vector, subtracts them to obtain the initial control error, and the virtual PID controller calculates the first control signal based on the initial temperature error and the injected parameters (12.5, 0.8, 3.2). This control signal is combined with other relevant features in the initial operating condition feature vector, such as the outdoor temperature, to form a complete input vector. This vector is fed into the trained system identification model. The model performs one forward propagation to predict the device power consumption and indoor average temperature at the next time step t=5 minutes. The program then records the control error at t=0 minutes and the predicted power consumption at t=5 minutes. In the second step (t=5 minutes to t=10 minutes), the state of the simulation environment is updated to the state at t=5 minutes. Specifically, the indoor average temperature in the operating condition feature vector is updated to the new temperature value predicted by the model at the previous time step, and the device power consumption and its related hysteresis features are also updated accordingly. Based on this updated operating condition vector, the program extracts the new indoor average temperature and the fixed set temperature again, and calculates the new control error. The virtual PID controller calculates a second control signal based on this new error. This signal is combined with the updated operating condition vector again and input into the system identification model. The model predicts the equipment power consumption and indoor temperature at t=10 minutes. The program records the control error e(1) at t=5 minutes and the predicted power consumption at t=10 minutes. This process will continue 24 times until the simulation of 2 hours of operation is completed. After the simulation ends, the program has collected and stored two complete sets of data.Finally, the program integrates all the data recorded during the simulation to form two ordered time series: the first is the control error time series, which contains 24 values ​​arranged chronologically, representing the control error value at the start of each time step from t=0 to t=115 minutes (the beginning of the 24th step); the second is the device power consumption time series, also containing 24 values ​​arranged chronologically, representing the predicted device power consumption value at the end of each time step from t=5 minutes to t=120 minutes (the end of the 24th step). These two time series constitute the complete simulation results for the current candidate PID parameter set. The entire injection and simulation process is repeated for all 50 candidate PID parameter sets in the population, ultimately generating a corresponding pair of control error and device power consumption time series for each parameter set.

[0041] Specifically, when initializing candidate PID parameter sets and injecting them into the configured simulation environment, the simulated control error time series may suffer from time-series distribution distortion due to differences between candidate PID parameters and the actual optimal parameters, as well as differences between the virtual PID controller and the actual PID controller. Traditional performance evaluation functions, such as differential error, are sensitive to the overall magnitude of the error but insensitive to its distribution over time. For example, a control process with a fast response but a small overshoot may yield similar error values ​​to a slow-response but smooth, overshoot-free process. However, in high-performance PID control, the former represents superior control quality. The ideal control error time series distribution should approximate a rapidly decaying, oscillating, and overshoot-free ideal form.

[0042] Therefore, to compensate for this temporal distribution distortion and fundamentally improve control quality, a metric is needed to measure the morphological difference between two temporal distributions. Compared to other metrics such as KL divergence, the Wasserstein distance is more advantageous in this scenario because it provides a meaningful gradient even when the two distributions have no overlap, which is crucial in subsequent optimization. This step introduces the Wasserstein distance to compare the temporal distribution of the simulation error with a pre-defined target error distribution representing the ideal control form (e.g., a Gaussian distribution with a mean of 0 and a small variance). Through optimization, the simulation error sequence is actively corrected to approximate the ideal state in dynamic form, thereby guiding the optimization of PID parameters from simply reducing the error magnitude to a more advanced approach of shaping the ideal error form.

[0043] Furthermore, in one possible implementation, the correction of the control error time series based on the Wasserstein distance error distribution morphology to obtain the corrected control error time series includes: The control error time series and the initial virtual disturbance vector are combined at location points to obtain the initial corrected control error time series, i.e.:

[0044] in, It is a time series of control errors. It is a vector added according to its position points. It is the initial virtual perturbation vector. This is the control error time series after initial correction. It should be understood that directly optimizing the original control error time series is impractical, as it is determined by fixed PID parameters and the simulation environment. By introducing a virtual disturbance vector initially set to zero or a random small value, and adding it element-wise to the original error series, an initial corrected control error time series can be generated. This new series retains the basic information of the original error while becoming adjustable due to the presence of the disturbance vector. This establishes an entry point for an optimization framework, transforming a fixed optimization problem into a parameterized problem solvable by adjusting the disturbance vector. Specifically, It is a variable initial correction sequence used for subsequent evaluation. Meanwhile, for Apply constraints ,in It is a preset, very small positive number determined through empirical setting or cross-validation, such as 0.01. This ensures that the correction amount is always small, avoids system divergence, and guarantees the stability and precision of the optimization.

[0045] The initial corrected control error time series is sorted based on its eigenvalues ​​to obtain an initially corrected control error sorted time series. Correspondingly, the Wasserstein distance, when calculating its one-dimensional discrete form, relies on comparing the cumulative distribution functions or quantile functions of the two distributions. In practice, this is equivalent to sorting the sample points of the two distributions and comparing them one by one. Therefore, before calculating the distance, all eigenvalues ​​in the initially corrected control error time series need to be arranged in ascending order to transform the disordered time series data into an ordered sample distribution representation. This is a preprocessing step for calculating the discrete Wasserstein distance. This generates an initially corrected control error sorted time series that no longer reflects the change of error over time, but rather the distribution of error values. This makes direct comparison with similarly sorted ideal distribution samples possible, laying the foundation for quantifying the morphological differences between the two distributions.

[0046] The mean of the control error is obtained by calculating the mean distance between adjacent features in the sorted time series of the control error after the initial correction.

[0047] in, and It is the first in the time series of control error sorting after initial correction. The and the first After initial correction, control error. It is the number of bits in the time series sorted by the control error after the initial correction. This controls the mean error. It should be understandable that the formula calculates the average of the squared differences between adjacent elements in the sorted sequence. This reflects the intrinsic scale or density of the current error distribution. When calculating the Wasserstein distance, the difference between corresponding points is multiplied by this... This is equivalent to performing a normalization or scale alignment, making the distance calculation less sensitive to the absolute size of the data and more focused on its relative distribution. It provides a dynamic, data-driven scaling factor for Wasserstein distance calculation, making the distance metric more robust and enabling a fair comparison of the morphological differences between error sequences of different magnitudes and the ideal distribution, thus improving the stability and accuracy of subsequent optimization processes.

[0048] A Gaussian distribution is used to sample the control error from the mean to obtain a control error reference vector. Correspondingly, a gold standard is needed as a reference to measure the quality of the current error distribution. Therefore, firstly, an ideal target error distribution is defined, representing the ideal error shape expected to be seen in the real world. In this embodiment, this distribution is set to have a mean of 0 and a variance of... Very small Gaussian distribution ,For example This can be set to 0.05, which represents the desired outcome where the error signal is concentrated close to zero most of the time. Then, sampling is performed from this ideal Gaussian distribution to generate a vector of the same length as the error sequence (length L). This vector controls the error reference vector to create a concrete, comparable instance of the ideal error distribution. This provides a clear optimization objective for subsequent distance calculations: by adjusting the perturbation vector, the distribution of the corrected error sequence should approximate the distribution of this reference vector as closely as possible.

[0049] The Wasserstein distance is calculated based on eigenvalues ​​from the initially corrected control error sorted time series and the control error reference vector, i.e.:

[0050] in, Calculated for absolute value. It is the first in the control error reference vector 1 eigenvalue, This refers to the Wasserstein distance. As you can understand, it's calculated using the Wasserstein distance formula. This formula calculates the minimum cost required to transform the corrected error distribution into an ideal reference distribution. This cost intuitively reflects the morphological differences between the two distributions. In this way, the differences between two complex distributions are condensed into a single, optimizable scalar value: the Wasserstein distance. This provides a clear loss function for subsequent optimization algorithms. The smaller this distance value, the closer the corrected error sequence is to the ideal state in dynamics, and the higher the control quality.

[0051] The initial virtual perturbation vector is iteratively updated by applying gradient descent to the initially corrected control error sorting time series and the control error reference vector. The final virtual perturbation vector is obtained by minimizing the Wasserstein distance between the initially corrected control error sorting time series and the control error reference vector.

[0052] in, It is to find that and The virtual perturbation vector with the minimum Wasserstein distance between the elements is found. This step is the core of the entire correction process. It treats minimizing the Wasserstein distance as an optimization problem. Using optimization algorithms such as gradient descent, the gradient of the Wasserstein distance with respect to each element in the virtual perturbation vector is calculated, and the virtual perturbation vector is updated in small steps along the inverse direction of the gradient. In the forward computation of each iteration, when the initial corrected control error time series is sorted, this process not only generates a numerically ordered sequence but also records a crucial positional permutation map, which clarifies which time position each element in the sorted sequence originates from in the unsorted sequence. In the backpropagation phase, the gradient of the Wasserstein distance with respect to each element in the sorted sequence is first calculated. Subsequently, using the positional permutation map recorded in the forward computation, these gradient values ​​are inversely permuted. Specifically, the gradient value corresponding to the k-th position in the sorted sequence is accurately propagated back to the original time position of that element before sorting. In this way, the complete gradient of the loss function with respect to the unsorted initial corrected control error time series is reconstructed. Since this sequence is the element-wise sum of the original error sequence and the virtual perturbation vector, and the original error sequence remains constant in this iteration, this gradient can be directly used as the gradient of the loss function with respect to the virtual perturbation vector. Finally, using the gradient descent rule, a preset learning rate is determined experimentally, for example, a small value such as 0.001 to ensure convergence stability. This reconstructed gradient is then used to update each component of the virtual perturbation vector. This iterative process continues until the Wasserstein distance converges to its minimum or its change falls below a preset convergence threshold. The resulting virtual perturbation vector is the final virtual perturbation vector that optimizes the error distribution.

[0053] The corrected control error time series is obtained by applying the control error time series and the final virtual disturbance vector according to the location points, i.e.:

[0054] in, It is the final virtual perturbation vector. This is the corrected control error time series. In other words, it's the original, unmodified control error time series. The final virtual perturbation vector obtained in the previous step The abstract optimization result is obtained by performing element-wise addition. This is then visualized and applied to the original data. The final corrected control error time series is obtained. This new time series not only potentially reduces the overall error magnitude, but more importantly, its time series distribution (such as response speed, overshoot, and oscillation characteristics) has been optimized to closely approximate an ideal Gaussian distribution. This corrected series will serve as a better performance evaluation basis, guiding the final selection of PID parameters. Specifically, using this corrected series to calculate fitness can guide optimization processes such as genetic algorithms, prioritizing PID parameter sets that not only reduce error but also deliver superior dynamic response characteristics, thereby achieving control qualities with faster response, smaller overshoot, and greater stability.

[0055] The fifth stage is step 45, fitness score calculation. First, the composition of the performance index function needs to be clarified. This function is a mathematical expression designed to integrate considerations of the two core objectives: control accuracy and energy consumption. A PID controller should be able to quickly and stably control the temperature near the setpoint (i.e., small control error) while minimizing the energy consumption of the equipment. In this embodiment, the performance index function is designed as a weighted sum to balance these two sometimes conflicting objectives. Specifically, the performance index function is defined as: Fitness Score = we*(1 / (1+IAE))+wp*(1 / (1+TPC)), where IAE is the absolute error integral, calculated by summing the absolute values ​​of all values ​​in the control error time series. This index measures the cumulative degree of control deviation over the entire simulation period; a smaller value indicates more accurate and stable control. TPC is the total power consumption, calculated by summing all values ​​in the equipment power consumption time series. This index measures the total energy consumption over the entire simulation period; a smaller value indicates more energy-efficient control. `we` and `wp` are weighting coefficients, representing the importance placed on control accuracy and energy efficiency, respectively, and `we + wp = 1`. The setting of these two values ​​depends on the specific application scenario and optimization objective. For example, in an application where comfort is paramount, `we` can be set to 0.7 and `wp` to 0.3; while in an application where energy efficiency is the primary objective, `we` can be set to 0.4 and `wp` to 0.6. In this embodiment, `w_e` is set to 0.5 and `w_p` to 0.5, indicating equal emphasis on both. The use of the form 1 / (1 + ...) in the function transforms the minimization problem (minimizing error and power consumption) into a maximization problem (maximizing fitness score), which aligns with the conventional setting in genetic algorithms where higher fitness is better. Adding 1 avoids a denominator of zero. Next, the specific calculation of the fitness score will be performed. Taking the simulation results of a candidate PID parameter group as an example, its output control error time series is [1.5, 0.8, 0.2, -0.1, ...], and the device power consumption time series is [50.2, 48.9, 47.5, 47.1, ...]. Adding the absolute values ​​of all 24 values ​​in the control error series yields a total error value, for example, IAE = 25.4. Adding the 24 values ​​in the device power consumption series yields a total power consumption value, for example, TPC = 1150.8. Substituting into the performance index function: Substituting the values ​​of IAE and TPC, along with the preset weights, into the formula: Fitness Score = 0.5 * (1 / (1 + 25.4)) + 0.5 * (1 / (1 + 1150.8)), a specific value is obtained, for example, 0.0194. This value, 0.0194, is the fitness score obtained by this candidate PID parameter group in this round of evaluation. This calculation process is performed on all 50 candidate PID parameter groups in the population one by one, and finally each parameter group is assigned a specific fitness score.

[0056] The sixth stage, step 46, involves selecting the optimal PID parameters. The program iterates through and compares the data of the final population. It checks the fitness score associated with each individual, i.e., each candidate PID parameter set. The program maintains a temporary variable to store the currently discovered highest fitness score and its corresponding PID parameter set. For example, the program first reads the first parameter set and its fitness score ((Kp1, Ki1, Kd1), 0.0194). Then it reads the second parameter set and its score ((Kp2, Ki2, Kd2), 0.0251). Since 0.0251 is greater than 0.0194, the program updates the highest score to 0.0251 and records the corresponding parameter set (Kp2, Ki2, Kd2). This comparison process continues until all 50 individuals in the population have been traversed. After the traversal, the program determines the unique individual in the entire population with the highest fitness score. For example, the 42nd candidate PID parameter set (Kp=8.7, Ki=0.45, Kd=1.2) was found to have the highest fitness score of 0.0312. The program then extracted this PID parameter set (8.7, 0.45, 1.2). This extracted parameter set is formally defined as the convergent optimal PID parameters for the current operating condition feature vector.

[0057] Finally, in step 47, the optimal PID parameters are selected. The program creates a new data structure, such as a data pair or a row in a table, to store the question-answer pair. This data structure contains two logical parts: the first part is the input description of the operating conditions, i.e., the current operating condition feature vector; the second part is the output decision result, i.e., the converged optimal PID parameters. Continuing the previous example, the current operating condition feature vector, which serves as the starting point for optimization, is a high-dimensional vector, such as the k-th cluster center Ck. After convergence by the genetic algorithm, the converged optimal PID parameters are a three-dimensional vector (8.7, 0.45, 1.2). The operation of constructing data pairs pairs these two vectors. In the program implementation, this can be represented by creating a data pair: (C_k, (8.7, 0.45, 1.2)). This data pair is defined as an optimal policy data in the optimal policy dataset. It clearly expresses the following mapping relationship: when the air conditioning operating condition is similar to the characteristics described by Ck, the optimal PID control parameters should be set to (8.7, 0.45, 1.2). This construction process is the last step in the optimization loop for each representative operating condition. Once the optimization process (steps 42 to 46) for a current operating condition feature vector is complete, this step is executed once, generating a corresponding optimal strategy data. Subsequently, the entire process restarts with a new sampled operating condition vector as the current operating condition feature vector, eventually generating another optimal strategy data. After all K, for example, 100 sampled operating condition vectors have completed their respective optimizations and generated corresponding optimal strategy data, the program collects all these generated data pairs, forming a set containing 100 records. This set is the optimal strategy dataset, where each record is a mapping between an operating condition and an optimal strategy.

[0058] In step 5, a lightweight supervised learning model is trained on the optimal policy dataset in the cloud to obtain a proxy policy model, which is then deployed at the edge. Correspondingly, the preceding steps, through complex offline simulation optimization, have constructed an expert knowledge base containing a series of typical operating conditions and their corresponding optimal PID parameters. However, this dataset itself consists of discrete and finite question-answer pairs, which cannot be directly used for real-time, continuously changing field conditions. More importantly, the simulation optimization process in the cloud is computationally intensive and time-consuming, completely failing to meet the stringent requirements of industrial controllers for low latency and high reliability. To transform the powerful analysis and optimization capabilities of the cloud into a practical tool capable of rapid response in the industrial field, a method is needed to compress and generalize this expert knowledge base. Therefore, this application trains a lightweight proxy model to learn and mimic the expert decision-making logic in the optimal policy dataset, generating a computationally efficient, compact model capable of generalizing from discrete knowledge points to any new operating condition, and deploying it at the edge, thereby achieving intelligent and adaptive adjustment of field controller parameters while meeting real-time requirements.

[0059] In one possible implementation, step 5 is performed as follows: In one possible implementation, the lightweight supervised learning model is a gradient boosting tree (XGBoost) model. It should be understood that gradient boosting (XGBoost) is an efficient, flexible, and scalable ensemble learning algorithm that makes predictions by iteratively constructing a series of decision trees. That is, each newly generated tree is designed to fit the residual (i.e., the difference between the true value and the predicted value) of the predictions from all trees in the previous round. The XGBoost model itself is not a single structure, but an additive model composed of multiple decision trees (weak learners). Its final prediction result is the sum of the predictions from all trees. The training process of the model involves learning the structure of each tree (split nodes, split thresholds) and the weight values ​​of each leaf node. The training process is as follows: The optimal policy dataset is processed. The current working condition feature vector in the dataset is used as the input feature (X) of the model, and the optimal PID parameters (Kp, Ki, Kd) after convergence are used as the output label (Y) of the model. Since XGBoost can handle multi-output regression tasks, three independent XGBoost models can be trained for the three parameters Kp, Ki, and Kd respectively. In this embodiment, the latter is adopted, i.e., three models are trained separately: model Kp, model Ki, and model Kd. Taking the training of model Kp as an example, the input X is the working condition feature vector, and the output Y is the corresponding Kp value. Before training, the hyperparameters of the XGBoost model need to be set. These hyperparameters determine the complexity of the model and the learning process. Key hyperparameters include: the number of trees is set to 200, indicating that the model consists of 200 decision trees. The maximum depth of each tree is set to 5 to prevent the model from becoming too complex and overfitting. The learning rate or step size is set to 0.1 to control the contribution of each tree to the final result; a smaller value makes the model more robust. Other regularization parameters, such as minimum split gain, L1 regularization term weights, and L2 regularization term weights, are also appropriately set to control the model complexity. It is worth mentioning that these hyperparameters were ultimately determined on the validation set with the goal of maximizing model performance metrics through automated tuning methods such as grid search, random search, or Bayesian optimization.

[0060] Taking model Kp as an example, the training process is iterative. First, the model starts with a simple initial prediction, such as the average of all training sample Kp values. Then, in the first iteration, the residual between the true Kp value and the current predicted value for each training sample is calculated. The program builds the first decision tree, whose learning objective is to predict these residuals as accurately as possible. The tree construction process uses a greedy algorithm to find the optimal split point to maximize information gain. After the first tree is built, its prediction (multiplied by the learning rate) is added to the initial prediction to obtain a new prediction. In the second iteration, the residual between the true Kp value and the new predicted value is calculated, and a second tree is built to fit these new residuals. This process continues until 200 trees are built. Each new tree corrects the accumulated error of all previous trees. The final model Kp internally incorporates the structure and leaf node weights of these 200 decision trees. When a new feature vector of a working condition is input, the vector will reach a leaf node along the splitting path in each tree, obtaining a leaf node weight value. The final Kp prediction value is the weighted sum of these 200 weight values. The same process is used to train models Ki and Kd. After training, three independent, trained XGBoost models are obtained. These three models together constitute the surrogate policy model.

[0061] The second phase involves deploying the agent strategy model at the edge. This means that industrial control scenarios have extremely high real-time requirements, with control decisions needing to be made within milliseconds or seconds. If every parameter adjustment requires a request to the cloud, network latency would be unacceptable, and a network outage would paralyze the entire intelligent control function, compromising reliability. Once the XGBoost model is trained, its inference process (i.e., prediction process) is extremely fast, requiring far less computation than cloud-based simulation optimization, and can run efficiently on edge devices with limited computing power.

[0062] The deployment process includes: First, saving the three pre-trained XGBoost models in the cloud as standard format files (such as PMML or ONNX), or directly as a format loadable by a specific programming language library (such as Python's pickle file). Next, these three model files are downloaded from the cloud server to the target edge controller via a secure network channel. Then, an inference engine for loading and executing the XGBoost models is integrated into the edge controller's application. Once the edge controller collects real-time sensor data and constructs a condition feature vector, it no longer needs to request data from the cloud; instead, it directly calls the three agent policy models locally.

[0063] In step 6, real-time sensor data is acquired at the edge. It is understood that the preceding steps have successfully condensed the complex optimization results from the cloud into a lightweight, efficient agent strategy model and deployed it to the edge controller in the industrial field. This model has the ability to output optimal PID parameters in real time based on operating conditions. However, the model itself is only a static decision-making tool; it requires dynamic input reflecting the current state of the real physical world to make meaningful and timely judgments. To enable the agent strategy model to perceive the instantaneous changes in the air conditioning equipment and its environment, and thus make precise and adaptive adjustments to control parameters, it needs to be provided with the latest and most accurate field operating data. Therefore, this application directly and periodically collects real-time readings from field sensors at the edge.

[0064] In one possible implementation, step 6 is performed as follows: First, the content of the real-time sensor data needs to be defined. This data is completely consistent in type and format with the historical data used for training in the cloud, because the input feature dimensions of the agent policy model are fixed. These data fields form the basis for describing the current instantaneous operating conditions, and their content may include, but is not limited to: environmental and indoor state quantities, such as outdoor temperature obtained by sensors deployed outside the building, and indoor average temperature calculated by averaging multiple indoor probes; control targets, such as the set temperature set by the operator on the control interface; and equipment status and controlled variables, such as instantaneous equipment power consumption data obtained by the power monitoring module.

[0065] Secondly, there's the data acquisition method. The edge controller communicates directly with the field sensor network or building automation system (BAS) through its integrated communication interface (such as industrial protocols like Modbus and BACnet). The edge controller proactively initiates a data acquisition request at a preset, fixed time period (e.g., every 5 minutes).

[0066] The specific implementation process is the same as acquiring historical sensor data streams in terms of data fields, but the executing entity and data source are different. Historical data is queried in batches from a cloud database, while real-time data is read instantly from the physical interface at the edge. For example, at a certain point in time, a timed task in the edge controller is triggered: the controller sends a read command to the outdoor temperature sensor to obtain the current outdoor temperature value, for example, 28.5 degrees Celsius. The controller polls multiple indoor temperature probes, reads their respective temperature values, and then performs a weighted average calculation locally to obtain the indoor average temperature, for example, 24.2 degrees Celsius. The controller reads the current set temperature from its internal storage or the host computer interface, for example, 24.0 degrees Celsius. The controller sends a read command to the smart meter connected to the chiller unit or main distribution cabinet to obtain the current instantaneous total power consumption, for example, 150 kilowatts.

[0067] Finally, data integration is performed. The edge controller binds all data values ​​acquired within the acquisition cycle with the current timestamp to form a data record. For example, at timestamp 2023-07-15 14:05:00, the generated data record is (timestamp: "2023-07-15 14:05:00", outdoor temperature: 28.5, indoor average temperature: 24.2, set temperature: 24.0, device power consumption: 150.0). This single data record with a timestamp is the real-time sensor data.

[0068] In step 7, at the edge, the real-time sensor data is input into the proxy strategy model to obtain updated PID parameters. That is, the original sensor data itself cannot be directly used by the model, and the model itself is merely a standby tool. To truly achieve adaptive adjustment of the control strategy, these two aspects must be combined to complete the crucial leap from perception to decision-making. Therefore, inputting real-time data into the proxy strategy model for inference calculation can activate and utilize the expert knowledge embedded in the model, transforming abstract real-time operating conditions into specific, optimized control commands (i.e., PID parameters), thereby providing the field controller with immediate and accurate adjustment basis.

[0069] In one possible implementation, step 7 is performed as follows: First, real-time feature engineering. The input dimensions and feature definitions of the proxy strategy model are determined during cloud training. Therefore, the edge device needs to process the acquired raw real-time sensor data into a real-time condition feature vector that is completely consistent with that used during training. This process is the same as the feature construction stage in cloud historical data processing, but the object of the operation is a single real-time data record.

[0070] Continuing the previous example, the real-time data record obtained by the edge controller is as follows: Timestamp: "2023-07-15 14:05:00", Outdoor temperature: 28.5°C, Indoor average temperature: 24.2°C, Set temperature: 24.0°C, Device power consumption: 150.0 kW. Simultaneously, the controller maintains a short-term historical data queue for calculating time-related derived features. The first step is to calculate the temperature error, i.e., the difference between the indoor average temperature and the set temperature: 24.2 - 24.0 = 0.2 degrees Celsius. At the same time, the building heat load is calculated, following the same formula as in historical data processing, such as C * (28.5 - 24.2) = 5.16 degrees Celsius, where C is 1.2. The second step involves the controller retrieving the device power consumption values ​​from the previous few time points (e.g., t-1, t-2, t-3) from its internal historical data queue to construct multi-order device power consumption hysteresis features. For example, power consumption (t-1) = 148.0, power consumption (t-2) = 147.5. The third step is to calculate the average power consumption and power consumption change rate of the device based on the data in the historical queue. The fourth step is to arrange the core fields in the original data and all the calculated derived features in the exact same order as during model training to form a real-time operating condition feature vector. For example, this vector might be [28.5, 24.2, 24.0, 150.0, 0.2, 5.16, 148.0, 147.5, ...].

[0071] The second stage is model inference. The real-time operating condition feature vector is input and fed into a locally deployed proxy policy model for computation. This proxy policy model consists of three independent, pre-trained Gradient Boosting Tree (XGBoost) models, corresponding to the prediction of parameters Kp, Ki, and Kd, respectively. The real-time operating condition feature vector is simultaneously input into these three models. Taking model Kp as an example, it internally contains a complex structure of 200 decision trees. When the feature vector is input, it propagates through each tree: starting from the root node, it compares the feature values ​​of a specific dimension of the vector with the node's set split threshold to determine whether to move to the left or right child node. This process proceeds layer by layer downwards until a leaf node is reached. Each leaf node stores a pre-learned weight value. The final output of model Kp is the weighted sum of the weight values ​​of the 200 leaf nodes obtained after traversing the 200 trees. This process does not involve any iteration or optimization; it is merely a series of predetermined path searches and addition operations, thus the computational overhead is extremely small and the speed is extremely fast. Similarly, models Ki and Kd perform the exact same inference process on the same input feature vector, each outputting a precise prediction. The program collects the outputs of these three models. For example, model Kp outputs 8.7, model Ki outputs 0.45, and model Kd outputs 1.2. The program combines these three values ​​into a triple (8.7, 0.45, 1.2).

[0072] This triplet (8.7, 0.45, 1.2) represents the updated PID parameters, indicating the optimal control strategy determined by the agent strategy model based on the current real-time operating conditions. These updated PID parameters are immediately sent to the underlying PID controller in the central air conditioning system that actually performs the control task. This controller updates its internal Kp, Ki, and Kd parameters to these new values. Starting from the next control cycle, the PID controller adjusts the actuators (such as chilled water valve opening and pump frequency) according to these new, more suitable parameters, thereby achieving precise and energy-efficient control of the air conditioning operation. This complete closed loop, from data acquisition to parameter update, repeats continuously at 5-minute intervals to ensure the control strategy always matches the changing operating conditions.

[0073] In summary, the integrated controller intelligent analysis and management method based on the embodiments of this application is explained, which lies in separating and deploying computationally intensive tasks from tasks with high real-time requirements. Specifically, in the cloud with abundant computing power, historical data is used to complete the training of a complex air conditioning system identification model and the offline optimization of the optimal PID parameter strategy, generating an expert strategy library containing various operating conditions and their corresponding optimal control parameters. Subsequently, a lightweight proxy model with low computational cost and fast inference speed is trained based on this strategy library. This proxy model is deployed on a resource-constrained edge controller. In this way, the edge does not need to perform complex online optimization calculations; it only needs to input real-time operating condition data into the proxy model to quickly infer the current optimal PID parameters. This cloud-based training and edge-based inference architecture cleverly solves the core contradiction mentioned in the background technology between complex AI algorithms and the low computing power and high real-time requirements of field controllers, realizing online, real-time, and intelligent self-tuning of controller parameters, effectively bridging the gap between advanced algorithms and industrial applications.

[0074] Figure 5 This is a block diagram of an integrated controller intelligent analysis and management system according to an embodiment of this application. Figure 5As shown, the integrated controller intelligent analysis and management system 100 according to an embodiment of this application includes: a historical sensor data stream acquisition module 110, used to acquire historical sensor data streams of an air conditioning system in the cloud; a working condition feature engineering module 120, used to perform working condition feature engineering on the historical sensor data streams of the air conditioning system in the cloud to obtain an air conditioning working condition feature engineering dataset; a system identification model training module 130, used to train a system identification model in the cloud based on the air conditioning working condition feature engineering dataset to obtain a trained system identification model; an optimal strategy dataset generation module 140, used to generate optimal PID controller parameters in the cloud based on the trained system identification model to obtain an optimal strategy dataset; a proxy strategy model training module 150, used to train a lightweight supervised learning model in the cloud based on the optimal strategy dataset to obtain a proxy strategy model, wherein the proxy strategy model is deployed at the edge; a real-time sensor data acquisition module 160, used to acquire real-time sensor data at the edge; and a PID parameter update module 170, used to input the real-time sensor data into the proxy strategy model at the edge to obtain updated PID parameters.

[0075] Here, those skilled in the art will understand that the specific operations of each step in the above-described integrated controller intelligent analysis and management system have been referenced above. Figures 1 to 4 The integrated controller's intelligent analysis and management method has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. An intelligent analysis and management method for an integrated controller, characterized in that, include: In the cloud, acquire historical sensor data streams from the air conditioning system; In the cloud, the historical sensor data stream of the air conditioning system is subjected to operating condition feature engineering to obtain an air conditioning operating condition feature engineering dataset. In the cloud, a system identification model is trained based on the air conditioning operating condition feature engineering dataset to obtain a trained system identification model. In the cloud, optimal PID controller parameters are generated based on the trained system identification model to obtain the optimal policy dataset, including: Extract the current operating condition feature vector from the air conditioning operating condition feature engineering dataset; Based on the trained system identification model and virtual PID controller, a configured simulation environment is constructed; Initialize a population containing multiple candidate PID parameter sets; The candidate PID parameter set is injected into the configured simulation environment to obtain the control error time series and the device power consumption time series; The method involves correcting the control error time series based on Wasserstein distance to obtain a corrected control error time series. This includes: combining the control error time series and the initial virtual perturbation vector by location points to obtain an initial corrected control error time series; sorting the initial corrected control error time series based on eigenvalues ​​to obtain an initial corrected control error sorted time series; calculating the mean distance between adjacent eigenvalues ​​in the initial corrected control error sorted time series to obtain the control error mean; sampling the control error using a Gaussian distribution based on the control error mean to obtain a control error reference vector; calculating the Wasserstein distance between the initial corrected control error sorted time series and the control error reference vector based on eigenvalues ​​to obtain the Wasserstein distance; iteratively updating the initial virtual perturbation vector using gradient descent on the initial corrected control error sorted time series and the control error reference vector, where the final virtual perturbation vector is obtained if the Wasserstein distance between the initial corrected control error sorted time series and the control error reference vector is minimized; and combining the control error time series and the final virtual perturbation vector by location points to obtain the corrected control error time series. The corrected control error time series and the device power consumption time series are input into the performance index function to obtain the fitness score of the candidate PID parameter group. Select the candidate PID parameter group corresponding to the highest fitness score as the optimal PID parameter after convergence of the current working condition feature vector; The current operating condition feature vector and the converged optimal PID parameters are constructed into a data pair to obtain an optimal strategy data in the optimal strategy dataset; In the cloud, a lightweight supervised learning model is trained on the optimal policy dataset to obtain an agent policy model, which is then deployed at the edge. At the edge, acquire real-time sensor data; At the edge, the real-time sensor data is input into the agent strategy model to obtain updated PID parameters.

2. The integrated controller intelligent analysis and management method according to claim 1, characterized in that, In the cloud, the historical sensor data stream of the air conditioning system is subjected to operating condition feature engineering to obtain an air conditioning operating condition feature engineering dataset, including: The historical sensor data stream of the air conditioning system is preprocessed to obtain a cleaned and aligned time series table; The cleaned and aligned time series table is constructed and vectorized with working condition features to obtain multiple feature columns, and the multiple feature columns are merged with the cleaned and aligned time series table to obtain an enhanced feature table. The feature row vector at time point t in the enhanced feature table is used as the i-th training sample; The label of the i-th training sample is a key controlled variable in the enhanced feature table at time point t+1. The labels of the i-th training sample and the i-th training sample are merged to obtain a training sample in the air conditioning operating condition feature engineering dataset.

3. The integrated controller intelligent analysis and management method according to claim 2, characterized in that, The features in the multiple feature columns include temperature error, building heat load, time lag characteristics of equipment power consumption, average equipment power consumption, and equipment power consumption change rate.

4. The integrated controller intelligent analysis and management method according to claim 2, characterized in that, The key controlled variable is the device power consumption.

5. The integrated controller intelligent analysis and management method according to claim 1, characterized in that, Extracting the current operating condition feature vector from the air conditioning operating condition feature engineering dataset includes: K-Means clustering analysis was performed on the air conditioning operating condition feature engineering dataset to obtain K sample operating condition vectors; Extract any one of the K sampled operating condition vectors as the current operating condition feature vector.

6. The integrated controller intelligent analysis and management method according to claim 1, characterized in that, The lightweight supervised learning model is the Gradient Boosting Tree (XGBoost) model.

7. An integrated controller intelligent analysis and management system, used to execute the integrated controller intelligent analysis and management method according to any one of claims 1-6, characterized in that, include: The historical sensor data stream acquisition module is used to acquire historical sensor data streams of the air conditioning system in the cloud. The operating condition feature engineering module is used to perform operating condition feature engineering on the historical sensor data stream of the air conditioning system in the cloud to obtain an air conditioning operating condition feature engineering dataset. The system identification model training module is used to train the system identification model in the cloud based on the air conditioning operating condition feature engineering dataset to obtain the trained system identification model. The optimal strategy dataset generation module is used to generate optimal PID controller parameters in the cloud based on the trained system identification model to obtain the optimal strategy dataset. The agent strategy model training module is used to train a lightweight supervised learning model on the cloud based on the optimal policy dataset to obtain an agent strategy model, which is deployed at the edge. The real-time sensor data acquisition module is used to acquire real-time sensor data at the edge. The PID parameter update module is used to input the real-time sensor data into the agent strategy model at the edge to obtain the updated PID parameters.

Images