A federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system

Abstract

The application relates to the technical field of deep learning, in particular to a federal learning collaborative modeling method for energy efficiency optimization of a central air conditioning system, which comprises the following steps: acquiring system operation state data, calculating real-time flow data, combining the real-time flow data with supply and return water temperatures, and calculating real-time load values; inputting environmental data into a benchmark generation model to obtain a load mode; inputting system operation data into an energy efficiency model to generate a dynamic remodeling coefficient; through a parameter transformation network, taking the dynamic remodeling coefficient as a transformation parameter of the network to generate a load prediction curve; based on the load prediction curve, solving through an optimal control algorithm to generate a collaborative control instruction, adjusting the operation frequency through the control instruction; using the dynamic remodeling coefficient as a label to train the energy efficiency model and generate updated parameters; after a server aggregates parameters to generate a global optimization model, the global optimization model is distributed, and the energy efficiency model is iteratively updated. Through the federal learning collaborative modeling, the system energy efficiency optimization is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system. BACKGROUND

[0002] The intelligent level of control and management of complex mechatronic systems such as central air conditioning systems in modern industry and construction fields is directly related to the operation efficiency and cost of the system. Such systems are characterized by multivariable, strong coupling, nonlinearity and dynamic time-varying, which puts high requirements on the accuracy and adaptability of the control strategy. Therefore, it is of great technical value to construct an intelligent control method that can deeply understand and optimize the running state of such systems.

[0003] For the control optimization of such complex systems, model-based predictive control is one of the current mainstream technical paths. However, in practice, the predictive model constructed based on the data of a single system still faces the challenge of generalization ability. First, the running state of a complex system is multi-dimensional and high-order, and is affected by many internal and external factors in a nonlinear manner. This leads to the model trained only relying on the limited historical data of a single system, when facing rare working conditions or sudden changes in external environment that have not been learned, its prediction accuracy and robustness are difficult to guarantee, which further affects the optimality of control decisions.

[0004] Secondly, in many application scenarios, there are a large number of similar systems with similar physical structures but different running modes. Under the existing technical architecture, the control model of each system is usually trained and deployed independently, forming a data island. This mode hinders the knowledge sharing and experience transfer between different systems, resulting in each model having to start from scratch. Especially for newly deployed systems with insufficient data accumulation, the model needs a long cold start period to reach the ideal performance. At the same time, this also makes any single model unable to obtain a more extensive global cognition beyond its own data boundary.

[0005] Therefore, how to break through the limitations of data islands while protecting the privacy of each system's data has become a technical problem to be solved in the field of intelligent control and data processing.

[0006] To this end, a federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system is proposed. SUMMARY

[0007] The purpose of the present application is to provide a federal learning collaborative modeling method for optimizing the energy efficiency of a central air conditioning system, which realizes system energy efficiency optimization through federal learning collaborative modeling. The data of the system running state is obtained, the real-time flow data is calculated, the real-time load value is calculated combined with the real-time flow data and the supply and return water temperature; the environmental data is input into the benchmark generation model to obtain the load pattern; the system running data is input into the energy efficiency model to generate the dynamic remodeling coefficient; the dynamic remodeling coefficient is taken as the transformation parameter of the network through the parameter transformation network to generate the load prediction curve; based on the load prediction curve, the optimal control algorithm is solved to generate the collaborative control instruction, and the running frequency is adjusted through the control instruction; the dynamic remodeling coefficient is used as a label to train the energy efficiency model and generate updated parameters; the server aggregates the parameters to generate a global optimization model for distribution, and iteratively updates the energy efficiency model.

[0008] To achieve the above purpose, the present application provides the following technical solutions:

[0009] A federal learning collaborative modeling method for optimizing the energy efficiency of a central air conditioning system, comprising:

[0010] The balance valve pressure difference data, the preset KV value data, the equipment running data, the supply and return water temperature and the outdoor environment data of the system running state are obtained; the balance valve pressure difference and the KV value are fused and calculated to obtain the real-time flow data, and the real-time flow data is combined with the supply and return water temperature to calculate the real-time load value;

[0011] The environmental data is input into the benchmark generation model to obtain the load pattern; the equipment running data, the environmental data, the real-time load value and the real-time flow data are input into the energy efficiency model to generate the dynamic remodeling coefficient; the dynamic remodeling coefficient is taken as the transformation parameter of the network through the parameter transformation network to execute time axis distortion and local amplitude modulation on the load pattern to generate the load prediction curve; based on the load prediction curve, the optimal control algorithm is solved to generate the collaborative control instruction, and the running frequency of the variable frequency water pump is adjusted through the control instruction;

[0012] In the local client, the dynamic remodeling coefficient is used as a label to train the energy efficiency model and generate updated parameters; the server aggregates the parameters to generate a global optimization model for distribution, and iteratively updates the energy efficiency model.

[0013] Preferably, the acquisition system runs the balanced valve pressure difference data, the preset KV value data, the equipment running data, the supply and return water temperature and the environment data, the equipment running data includes the running frequency and real-time power of the variable frequency water pump, and the environment data includes indoor and outdoor temperature and humidity and space occupancy state; the Kalman filtering algorithm is used for denoising and smoothing processing for the balanced valve pressure difference data and the real-time power; the space occupancy state is converted into a numerical characteristic through one-hot encoding; the preset KV value data, the running frequency, the supply and return water temperature and the indoor and outdoor temperature and humidity are normalized to eliminate the influence of different physical dimensions; the preprocessed data stream is synchronized and aligned based on a unified timestamp, and the time series interpolation method is used to fill in the missing values in the sequence.

[0014] Preferably, the process of calculating the real-time load value includes: substituting the balanced valve pressure difference data and the preset KV value into a valve flow coefficient calculation formula to calculate a theoretical flow reference value; dynamically adjusting internal correction parameters according to the real-time theoretical and actual flow deviation through an online adaptive corrector to output a real-time correction value; and weighting and summing the theoretical flow reference value and the real-time correction value to generate the real-time flow data during real-time calculation.

[0015] According to the supply and return water temperature data, the average temperature of the current fluid medium is calculated; the predefined discrete temperature nodes and the corresponding standard fluid density and specific heat capacity parameters are constructed into a reference state vector, and an interpolation weight vector is generated according to the position of the average temperature between the discrete temperature nodes; the interpolation weight vector and the reference state vector are subjected to vector dot product operation to solve the corresponding fluid density and specific heat capacity at the current average temperature; and the real-time flow data, the fluid density and the specific heat capacity and the supply and return water temperature difference are comprehensively considered to calculate and output the real-time load value through a thermodynamic energy balance equation.

[0016] Preferably, the process of inputting the environment data into the reference generation model to obtain the load mode includes: dividing the environment data according to a period to form environment data segments as independent samples; inputting each environment data segment into a reference generation model, the reference generation model is an encoder composed of a long short-term memory network, and the environment data segment is compressed and extracted into a low-dimensional latent vector; the low-dimensional latent vectors are collected, K-means clustering algorithm is used for clustering analysis in the low-dimensional latent vector space to identify typical daily modes with statistical significance; and the centroid vector of each cluster is output as the load mode.

[0017] Preferably, the process of generating the dynamic remodeling coefficient comprises: aligning the equipment operation data, environmental data, real-time load value and real-time flow data on a time axis, and at each synchronized time sampling point, splicing the values of all data sources in a predetermined order to construct a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a pre-trained energy efficiency model based on a Transformer architecture to perform weighting and information fusion on the feature vector; and the energy efficiency model maps and outputs the dynamic remodeling coefficient through a fully connected output layer after fusing the high-dimensional features.

[0018] Preferably, the process of generating the load prediction curve comprises: extracting elements corresponding to a first preset index range in the dynamic remodeling coefficient as time domain parameters for controlling time axis nonlinear transformation, and extracting elements corresponding to a second preset index range as amplitude parameters for controlling local amplitude variation; inputting the time domain parameters into a time domain transformation module of the parameter transformation network to generate a one-dimensional sampling grid, and using a differentiable interpolation algorithm to resample the load pattern according to the one-dimensional sampling grid to generate an intermediate load pattern with a nonlinearly distorted time axis; inputting the amplitude parameters into an amplitude modulation module of the parameter transformation network to generate a multiplicative mask containing a local scaling factor and equal in length to the intermediate load pattern; and multiplying the multiplicative mask and the intermediate load pattern point by point to obtain the load prediction curve.

[0019] Preferably, the process of generating the cooperative control instruction comprises: taking the load prediction curve as a system demand constraint, minimizing the energy consumption of the variable frequency pump as an optimization objective, and constructing a multivariate dynamic optimization problem with constraints; in each control period, using a sequential quadratic programming algorithm to rollingly solve the dynamic optimization problem to obtain an optimal control sequence; and extracting the first control instruction in the optimal control sequence as the cooperative control instruction at the current time for output.

[0020] Preferably, the process of iteratively updating the energy efficiency model comprises: at each client, taking the dynamic reshaping coefficient as the true value label of supervised learning and taking the multi-dimensional feature vector as the input, training the energy efficiency model through multiple rounds of calculation of the loss function between the predicted value and the true value label and adoption of the back propagation and gradient descent algorithm, calculating and extracting the model weight update amount of the local copy relative to the global model before training after the local training is completed, and uploading the update amount as the local model update parameter to the central server; the central server, after collecting the model weight update amount from multiple clients, adopts the federated average algorithm to perform weighted average on multiple update amounts, calculates the aggregated global model update amount, and uses the aggregated global model update amount to update the maintained global optimization model, and distributes the updated global optimization model to each client to replace the original local energy efficiency model, completing the iterative update.

[0021] Compared with the prior art, the application has the following beneficial effects:

[0022] 1. By combining high-precision real-time state estimation with a two-stage load prediction model, the real-time load is first accurately calculated to provide a solid data foundation for prediction, and then a typical daily load benchmark is generated using the model and dynamically distorted and modulated according to real-time multi-dimensional features. This benchmark + deviation approach can accurately capture load variation patterns and real-time disturbances, greatly improving the accuracy and reliability of load prediction.

[0023] 2. By applying accurate load prediction curves to the optimal control algorithm, this method realizes a leap from traditional passive response to proactive active optimization. The system can foresee future demand and perform rolling solution with the goal of minimizing energy consumption, planning optimal device operation strategies in advance to avoid energy waste caused by lagging response, and can tap into system energy-saving potential to achieve global and full-period energy efficiency optimization.

[0024] 3. By introducing a federated learning collaborative modeling framework, each client's data does not need to be uploaded, only the model update amount is trained and shared locally, which fundamentally guarantees the privacy and security of user data, breaks down data silos, and enables the global model to draw on the strengths of others and continuously evolve. The final generated model has stronger generalization ability and robustness and can adapt to more diverse operating conditions. BRIEF DESCRIPTION OF DRAWINGS

[0025] Fig. 1 A federated learning collaborative modeling method flowchart for central air conditioning system energy efficiency optimization is provided for the embodiments of the application.

[0026] Fig. 2 A rolling solution schematic diagram is provided for the embodiments of the application.

[0027] Fig. 3 A client-server interaction schematic diagram provided for the embodiment of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0029] Please refer to Figs. 1 to 3 The present application provides a kind of central air conditioning system energy efficiency optimization's federal learning collaborative modeling method, technical scheme is as follows:

[0030] Embodiment one:

[0031] A kind of central air conditioning system energy efficiency optimization's federal learning collaborative modeling method, specific process is as shown in Fig. 1 Including:

[0032] Balancing valve differential pressure data, preset KV value data of system running state are acquired, equipment running data, supply and return water temperature and outdoor environment data;Balancing valve differential pressure and KV value are fused and calculated, and real-time flow data is obtained, in combination with real-time flow data and supply and return water temperature, real-time load value is calculated;

[0033] Environment data is input into benchmark generation model, and load mode is obtained;Equipment running data, environment data, real-time load value and real-time flow data are input into energy efficiency model, and dynamic remodeling coefficient is generated;Through parameter transformation network, dynamic remodeling coefficient is used as the transformation parameter of network, and time axis distortion and local amplitude modulation are executed to load mode, to generate load prediction curve;Based on load prediction curve, by optimal control algorithm, to minimize the target of electric energy consumption, collaborative control instruction is generated, and the frequency of variable frequency water pump is adjusted by control instruction;

[0034] In client local, dynamic remodeling coefficient is used as label, and energy efficiency model is trained and updated parameter is generated;Server aggregates parameter to generate global optimization model and distributes, and energy efficiency model is iteratively updated.

[0035] Further, the acquisition system runs the balance valve pressure difference data, the preset KV value data, the equipment running data, the supply and return water temperature and the environment data, the equipment running data includes the running frequency and the real-time power of the frequency conversion water pump, and the environment data includes the indoor and outdoor temperature and humidity and the space occupation state; the Kalman filtering algorithm is used for denoising and smoothing processing for the balance valve pressure difference data and the real-time power; the space occupation state is converted into a numerical characteristic by one-hot encoding; the preset KV value data, the running frequency, the supply and return water temperature and the indoor and outdoor temperature and humidity are normalized to eliminate the influence of different physical dimensions; the preprocessed data stream is synchronized and aligned based on a unified timestamp, and the missing values in the sequence are filled by using the time series interpolation method.

[0036] Specifically, a series of original data used to represent the system state is collected in real time and periodically, in the embodiment, the equipment running data is defined as including the real-time running frequency (unit: Hz) and the real-time consumed power (unit: kW) of the frequency conversion water pump, and the environment data includes the indoor and outdoor dry and wet bulb temperature (unit: ℃) and the space occupation state (for example, no one, few people, many people and other discrete states); after the original data is acquired, a series of targeted preprocessing operations are immediately performed on the edge computing device locally in the client; since the original balance valve pressure difference data and the real-time power data usually contain high-frequency noise due to the influence of electromagnetic interference and other factors in the data collection process, the Kalman filtering algorithm is used to filter the two time series data in real time to eliminate random disturbances and abnormal outliers, and generate a smooth data sequence that can better reflect the real physical process. For the space occupation state category feature, the computer model cannot be directly processed, therefore, the one-hot encoding method is used to convert it into a multi-dimensional, sparse binary (0 or 1) numerical feature vector. For example, the three states of “no one”, “few people” and “many people” are converted into [1, 0, 0], [0, 1, 0] and [0, 0, 1] respectively, so that they can be effectively used by the model; since there is a huge difference in the numerical range of different physical quantities (for example, the power value may be thousands, while the temperature value is only dozens), in order to eliminate the adverse effects of the dimensional difference on the model training, the minimum and maximum normalization method is used to scale the values of all numerical features, including the preset KV value data, the running frequency, the supply and return water temperature and the indoor and outdoor temperature and humidity, to the interval [0, 1]; finally, since there may be slight deviations in the collection frequency of different data sources or data may be lost due to network problems, the preprocessed data stream is strictly synchronized and aligned based on a unified timestamp that is accurate to the second, and the missing values found in the alignment process are filled by using the linear time series interpolation method.

[0037] Through this flow, raw data from multiple sources, heterogeneous, can be transformed into a high-quality, standardized feature matrix. The process not only effectively improves the purity and usability of the data through steps such as denoising, encoding and normalization, but also ensures the integrity and consistency of the data set through synchronous alignment and missing value filling. This provides a solid and reliable data foundation for the training and inference of all subsequent advanced models, and is a key prerequisite for ensuring the final performance and stability of the entire system.

[0038] Further, the process of calculating the real-time load value includes: substituting the balance valve pressure difference data and the preset KV value into a valve flow coefficient calculation formula to calculate a theoretical flow reference value; through an online adaptive corrector, according to the real-time theoretical and actual flow deviation, dynamically adjusting an internal correction parameter, outputting a real-time correction value; in real-time calculation, the theoretical flow reference value and the real-time correction value are weighted and summed to generate the real-time flow data;

[0039] According to the supply and return water temperature data, the average temperature of the current fluid medium is calculated; the predefined discrete temperature nodes and the corresponding standard fluid density and specific heat capacity parameters are constructed as a reference state vector, and an interpolation weight vector is generated according to the position of the average temperature between the discrete temperature nodes; by performing vector dot product operation on the interpolation weight vector and the reference state vector, the corresponding fluid density and specific heat capacity at the current average temperature are calculated; the real-time flow data, fluid density and specific heat capacity, and the supply and return water temperature difference are integrated, and the real-time load value is calculated and output through the thermodynamic energy balance equation.

[0040] Specifically, the embodiment is based on a known fluid dynamics principle that the flow through the valve is proportional to the square root of the pressure difference between the two ends of the valve, and the preset KV value representing the flow capacity of the valve is taken as the proportional coefficient, thereby calculating a theoretical flow reference value as the basis for subsequent correction.

[0041] However, due to the complexity of actual working conditions such as pipeline aging and valve corrosion, there is usually a systematic deviation between the theoretical value and the real flow. In order to accurately compensate for the deviation, the embodiment uses an online adaptive corrector. The corrector dynamically and continuously adjusts a weight parameter in its interior according to the real-time theoretical and actual flow deviation (for example, by comparing with a small number of real flow meters in the system as calibration points), and outputs a real-time correction value. Finally, in real-time calculation, the theoretical flow reference value and the real-time correction value are dynamically weighted and fused to generate a high-precision real-time flow data.

[0042] The online adaptive modifier is built-in recursive least square algorithm, the core function is to continuously calculate a dynamic correction coefficient, by comparing the theoretical flow reference value and the real flow meter reading in the system as the calibration point, the deviation between the two, recursive least square algorithm according to this deviation, automatically, continuously update the correction coefficient, so that it can reflect the actual working condition influence such as pipeline aging, valve state change, finally, the real-time theoretical flow reference value will be multiplied by this dynamic correction coefficient, so as to obtain a more realistic real-time flow data.

[0043] The online adaptive modifier uses recursive least square algorithm with forgetting factor to calibrate the theoretical model of flow meter online, the specific implementation steps are as follows: parameter initialization, when running for the first time, set the initial value of correction coefficient to 1.0, which means that the initial moment is trusted to theoretical calculation, set the initial value of covariance matrix to a larger value, for example, 1000, which represents high uncertainty to initial coefficient, set the forgetting factor to 0.99, which is used to gradually reduce the influence of old data in iteration, so that the algorithm can adapt to the slow change of the system; in each control period (for example, every minute), perform an update, first, calculate the prediction error with a value of 0.3 according to the real flow measurement value and the theoretical flow calculation value of the period, then, calculate the current gain value using the covariance matrix of the last moment and the current theoretical flow calculation, finally, update the correction coefficient according to the prediction error and the gain, and update the covariance matrix at the same time, to prepare for the next iteration.

[0044] Then, the physical properties of the current fluid medium (usually water) need to be determined, in this embodiment, the average temperature of the current fluid medium is calculated according to the real-time collected supply and return water temperature data, the system internally defines a reference state vector in advance, which stores the standard fluid density (ρ) and specific heat capacity (c) parameters at corresponding temperatures in the form of discrete temperature nodes (for example, 5℃, 10℃, 15℃...), according to the position of the currently calculated average temperature among these discrete nodes, the system will generate an interpolation weight vector in real time, by performing vector dot product operation on the weight vector and the reference state vector, the real-time fluid density and specific heat capacity corresponding to the current average temperature can be accurately calculated, finally, the real-time flow data generated in the last step, the fluid density and specific heat capacity just calculated and the real-time supply and return water temperature difference are integrated, and the final real-time load value is calculated and output through the thermodynamic energy balance equation.

[0045] Through the flow, the physical mechanism-based theoretical calculation can be combined with an online adaptive data correction process; the mechanism + data hybrid modeling method makes the finally generated real-time flow data and load value more accurate than a single method, not only has the stability of the theoretical model, but also can compensate the systematic deviation caused by the real factors such as equipment aging and working condition change in real time; the method has excellent dynamic adaptability through the online adaptive corrector and real-time physical property parameter interpolation solution, can continuously calibrate itself to adapt to the long-term change of system performance, and ensures the long-term effectiveness of the calculation accuracy.

[0046] Further, the process of inputting the environment data into the benchmark generation model to obtain the load mode comprises: segmenting the environment data according to a period to form environment data segments as independent samples; inputting each environment data segment into a benchmark generation model, the benchmark generation model being an encoder composed of a long short-term memory network, compressing and extracting the environment data segment into a low-dimensional latent vector; collecting the low-dimensional latent vectors, performing clustering analysis in the low-dimensional latent vector space by using a K-means clustering algorithm, and identifying typical daily modes with statistical significance; outputting the centroid vector of each clustering cluster as the load mode.

[0047] Specifically, the time series of environmental data after preprocessing, containing multi-dimensional information such as indoor and outdoor temperature and humidity, and space occupancy state, is divided according to a fixed 24-hour cycle. Each divided data segment constitutes an environment data segment that can completely represent the single-day environmental change trajectory as an independent sample; each of the environment data segments is input into a reference generation model. In this embodiment, the model is an encoder based on a long short-term memory network; the encoder part is composed of two layers of stacked bidirectional LSTM networks, the first layer contains 128 hidden units, and the second layer contains 64 hidden units. The input data is a sequence of environmental data representing a whole day, and the output is a 32-dimensional latent vector. The model obtains feature extraction capability through pre-training, and the training target is to enable it to reconstruct the original input according to the latent vector. The mean square error is used as the loss function during training, and the Adam optimizer is used to adjust the network parameters; the encoder can compress and extract a multi-dimensional environment data segment containing multiple time steps (for example, 24x60=1440 minute-level data points in a day) into a fixed-length low-dimensional latent vector that can capture its core dynamic characteristics, using its memory and processing capability for time series data; after processing all historical environment data segments through the encoder, the system obtains a set of a large number of low-dimensional latent vectors. In this embodiment, the K-Means clustering algorithm is used to perform unsupervised clustering analysis of these vectors in the feature space composed of the low-dimensional latent vectors. This algorithm can automatically divide days with similar change rules (i.e., vectors with similar distances in the latent space) into the same cluster, thereby identifying statistically significant typical day patterns. For example, K=2 can be preset to distinguish between weekday and non-weekday patterns; for each identified cluster, the centroid vector of all member vectors (i.e., the average of all vectors in the cluster) is calculated. To convert the 32-dimensional centroid vector into a usable time series, it needs to be input into a decoder network (composed of two layers of LSTM networks) paired with the encoder. The decoder is responsible for reconstructing the latent vector into a complete, 1440-time-step standard time series curve. This reconstructed time series curve is output as the entity of the load pattern, providing a quantifiable, standardized baseline template for subsequent dynamic remodeling.

[0048] Through this process, deep learning models can automatically learn and extract statistically representative typical day patterns from high-dimensional, continuous environmental data. Not only does this effectively reduce the dimensionality of complex time series data and abstract features, but it also provides a quantifiable, stable baseline template that eliminates single-day random disturbances for subsequent load prediction, significantly improving the robustness and accuracy of the prediction system.

[0049] Further, the process of generating the dynamic remodeling coefficient comprises: aligning the device operation data, environment data, real-time load value and real-time flow data on a time axis, and at each synchronized time sampling point, splicing the values of all data sources in a predetermined order to construct a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a pre-trained energy efficiency model based on the Transformer architecture to perform weighting and information fusion on the feature vector; and the energy efficiency model maps and outputs the dynamic remodeling coefficient through a fully connected output layer after fusing the high-dimensional features.

[0050] Specifically, the real-time data streams from multiple sources, including device operation data, environment data, real-time load value and real-time flow data, are strictly synchronized and aligned on a time axis, and at each synchronized time sampling point, the values of all data sources at this moment are spliced in a predetermined and fixed order. In this embodiment, the real-time load value, real-time flow data, operation frequency and real-time power in the device operation data, and indoor and outdoor temperature and humidity in the environment data and the space occupation state vector after one-hot encoding are spliced in sequence, thereby constructing a multi-dimensional feature vector that can represent the instantaneous system state in a single vector form. Subsequently, the multi-dimensional feature vector is input into a pre-trained energy efficiency model based on the Transformer architecture. The core advantage of this model lies in its internal self-attention mechanism, which can automatically calculate and assign different weights to different elements (i.e., different physical quantities) in the feature vector, thereby effectively capturing the complex and nonlinear interdependent relationships between them and performing deep information fusion. Finally, the energy efficiency model maps and outputs the high-dimensional feature information fused by the self-attention mechanism into a set of fixed-dimensional parameterized vectors as the dynamic remodeling coefficient through a fully connected output layer. The vector quantitatively represents the deviation degree and direction of the current system real-time state relative to the load pattern, providing accurate driving parameters for subsequent dynamic transformation of the load pattern.

[0051] The energy efficiency model adopts the encoder architecture of Transformer, and the key configuration is that it is stacked by 4 encoder layers, each encoder layer uses 8 self-attention mechanisms, and the internal feedforward network dimension is 512. The input multi-dimensional feature vector is mapped to the internal working dimension of the model through a linear embedding layer, and position encoding is added to preserve the time order information of the data.

[0052] The process of the input energy efficiency model also includes an attention guiding process based on input data reliability evaluation, which enhances the robustness of the model when facing data anomalies. Before inputting the multi-dimensional feature vector into the energy efficiency model, each data input item in the vector is evaluated for reliability. The reliability score is calculated through range check and rate of change anomaly detection. The score is used to generate an attention guiding mask, which takes effect in the self-attention mechanism layer inside the energy efficiency model, reducing the weight of data features with low reliability scores in the information fusion process. For example, if a key data input item (such as the return water temperature reading) used to calculate the real-time load value is abnormal, its value frequently jumps in a short period of time, the reliability evaluation module will immediately reduce the score of this data item to a limited value (such as 0.1). When the feature vector containing this abnormal data enters the energy efficiency model, the attention guiding mask will make the model almost completely ignore the influence of this abnormal data, and instead rely more on other data items for comprehensive judgment, so that it can still output smooth and reasonable dynamic remodeling coefficients, avoiding the risk of performance degradation of the entire prediction and control system due to local data quality problems.

[0053] The specific implementation steps of the reliability evaluation algorithm include pre-setting a physically reasonable value range and a maximum allowed change rate for each data item. For example, the reasonable range of water supply temperature is set to [5, 10] degrees Celsius, and the maximum allowed change rate is set to 0.5 degrees Celsius per minute. It is determined whether the currently collected water supply temperature value (e.g., 12 degrees Celsius) is within the range of [5, 10] degrees Celsius. Since 12 degrees Celsius is outside the range, the range score is 0.0. The change rate of the current value is calculated. If the temperature was 7.2 degrees Celsius last minute and 7.1 degrees Celsius currently, the absolute value of the change rate is 0.1 degrees Celsius per minute, which is less than the maximum allowed change rate of 0.5 degrees Celsius, so the change rate score is 1.0. The final reliability score of the data point is equal to the product of the range score and the change rate score. For the abnormal value of 12 degrees Celsius, the final score is 0.0 multiplied by 1.0, resulting in 0.0 (unreliable). For the normal value of 7.1 degrees Celsius, the final score is 1.0 multiplied by 1.0, resulting in 1.0 (reliable). In self-attention calculation, the attention score of each feature is multiplied by its corresponding reliability score. Therefore, the attention of the data point of 12 degrees Celsius will be zeroed out, so it will be completely ignored by the model at the current time step.

[0054] By constructing multi-source real-time data into a unified feature vector and using the self-attention mechanism of the Transformer model, this process can automatically capture deep non-linear relationships between physical quantities. The dynamic remodeling coefficients output by the model represent the deviation of the current working condition from the baseline mode in a quantitative and interpretable manner, providing direct and reliable driving parameters for subsequent precise dynamic transformation.

[0055] Further, the process of generating the load prediction curve comprises: extracting elements corresponding to a first preset index range in the dynamic reshaping coefficient as a time domain parameter for controlling time axis nonlinear transformation, and extracting elements corresponding to a second preset index range as an amplitude parameter for controlling local amplitude variation; inputting the time domain parameter into a time domain transformation module of the parameter transformation network, generating a one-dimensional sampling grid, and using a differentiable interpolation algorithm to resample the load pattern according to the one-dimensional sampling grid, to generate an intermediate load pattern with a nonlinearly distorted time axis; inputting the amplitude parameter into an amplitude modulation module of the parameter transformation network, to generate a multiplicative mask containing a local scaling factor and being the same length as the intermediate load pattern; and multiplying the multiplicative mask and the intermediate load pattern point by point to obtain the load prediction curve.

[0056] Specifically, the generated dynamic reshaping coefficient vector is analyzed. The total dimension of the vector and the dimensions thereof assigned to time domain transformation and amplitude modulation are predefined as model hyperparameters in network architecture design, and the values thereof depend on the required control complexity of subsequent transformation. In this embodiment, the total dimension of the vector is defined as 16 dimensions, wherein the first part is defined as a time domain parameter for controlling time axis nonlinear transformation, and the first 8 elements of the vector are explicitly taken; and the second part is defined as an amplitude parameter for controlling local amplitude variation, and the last 8 elements of the vector are explicitly taken. Through this fixed rule based on model architecture, a unified coefficient vector is analyzed into two groups of independent parameters with explicit physical meaning for guiding different transformations.

[0057] Subsequently, the time domain parameter is input into a time domain transformation module of the parameter transformation network. In this embodiment, the module is a small feedforward neural network that transforms the input time domain parameter into a one-dimensional sampling grid defining a new time coordinate. Then, the system uses a differentiable linear interpolation algorithm to resample the load pattern according to the sampling grid. This process is equivalent to nonlinearly “stretching” or “compressing” the reference pattern on the time axis, thereby generating an intermediate load pattern with a distorted time axis; at the same time, the amplitude parameter is input into an amplitude modulation module of the parameter transformation network, which is also a small feedforward neural network that transforms the input amplitude parameter into a multiplicative mask containing a local scaling factor and being the same length as the intermediate load pattern. Each element in the mask vector corresponds to the amplitude scaling ratio of the intermediate load pattern at a certain time point.

[0058] Specifically, the 8 time-domain parameters outputted by the time-domain transformation module are used as a set of control points, and then a one-dimensional sampling grid with full-day time resolution is generated by cubic spline interpolation between these control points. Similarly, the amplitude parameter is also generated into a corresponding length of multiplicative mask through the interpolation method.

[0059] Finally, the multiplicative mask generated in the last step is multiplied point by point with the intermediate load pattern. Through this operation, the local amplitude of the intermediate load pattern is accurately adjusted, so that the static load pattern is dynamically and nonlinearly transformed and output as the final load prediction curve according to real-time data.

[0060] By decomposing the dynamic adjustment process into time axis distortion and local amplitude modulation, the process can finely reshape the load pattern in a structured manner, so that it can not only reflect the advance or delay of load peaks, but also simulate the increase and decrease of local intensity, significantly improving the fitting ability and expressiveness of the prediction curve to complex real-time working conditions.

[0061] Further, the process of generating the cooperative control instruction includes: taking the load prediction curve as a system demand constraint, minimizing the energy consumption of the variable frequency water pump as the optimization target, and constructing a multivariate dynamic optimization problem with constraints; in each control period, a sequential quadratic programming algorithm is used to solve the dynamic optimization problem, and the optimal control sequence is obtained; the first control instruction in the optimal control sequence is extracted as the cooperative control instruction at the current time for output. The specific rolling solution is shown in Fig. 2 .

[0062] Specifically, the present embodiment constructs the process as a dynamic optimization problem, the core of which is to predict the future dynamic behavior of the system using a mathematical model and perform rolling optimization within a limited time range. Within a certain future time period (e.g. 24 hours), an optimal water pump operation strategy is found to minimize the total power consumption of the variable frequency water pump. The optimization process adheres to a series of constraints: first, the physical laws of the system must be followed, which means that the optimization algorithm must take into account how the system state (such as pipe network pressure, water temperature) changes dynamically according to the control instruction (water pump frequency) and external disturbance (such as outdoor temperature); second, the cooling demand of the building must be met, and the system must produce cooling capacity greater than or equal to the load demand predicted in the previous step at any time within the future time period to ensure user comfort; third, the control instruction must be within the safe operating range of the equipment, for example, the operating frequency of the variable frequency water pump must always be within the minimum and maximum frequency set by the manufacturer; fourth, the key state parameters of the system must also be maintained within the preset range, including but not limited to the supply and return water temperature difference, the pressure at key points in the pipe network, etc., to ensure efficient and safe operation of the entire system. To solve this complex optimization problem with nonlinear characteristics, the present embodiment uses an efficient numerical optimization algorithm called sequential quadratic programming, which can effectively handle multivariate, multi-constrained nonlinear problems. Through a series of iterative calculations, the complex original problem is transformed into a set of easier-to-solve sub-problems, thus quickly and stably finding the optimal control strategy that satisfies all constraints. Finally, the control instruction is executed using a rolling optimization approach. At the beginning of each control period (e.g. every 15 minutes), the system will solve the optimization problem based on the latest real-time data and load prediction, obtaining an optimal control sequence covering the entire prediction period. However, the system does not execute the entire sequence, but only extracts and executes the first instruction in the sequence, i.e. the water pump frequency to be used at the current time, and enters the next control period. The system then obtains new data, updates the prediction, and performs another round of complete optimization calculation.

[0063] To construct the optimization problem, a data-driven surrogate model is first defined as a prediction model of the system behavior. The surrogate model is a three-layer fully connected neural network that establishes causal relationships between key system parameters by learning from historical data. The input is the water pump frequency (control input), and the output is the predicted value of the pipe network end pressure (system state) and the total energy consumption (optimization objective). At the same time, the constraints must be quantified to specify that the operating frequency of the variable frequency water pump must always be between 30Hz and 50Hz, and the cooling capacity produced by the system at any time must not be less than the demand value at the corresponding time on the load prediction curve.

[0064] The generating the cooperative control instruction further includes a device operation combination mode decision flow based on iterative optimization, enumerating all current feasible device operation combination modes; for each combination mode, a prediction model composed of a three-layer fully connected neural network and pre-trained according to historical device operation data and environment data is used for optimization solving; the optimization solving takes the current environment data as the fixed condition input of the model, takes minimizing the total energy consumption of the system as the goal, iteratively adjusts the device operation data (in this embodiment, the operation frequency of each variable frequency water pump and the operation frequency of the cooling tower fan) as the variable control input of the model, so that the total energy consumption of the system predicted by the model reaches the optimal operation frequency with the lowest total energy consumption, and the lowest total energy consumption value is recorded; after the evaluation of all feasible modes is completed, the lowest total energy consumption value recorded by each mode is compared, and the combination mode with the lowest total energy consumption and the optimal operation frequency corresponding thereto are selected as the cooperative control instruction and output. By comprehensively quantifying and optimizing all feasible device combinations, an efficient operation strategy is explored, so that the global optimum of system energy efficiency is found, and deep energy saving is realized.

[0065] Through this "rolling solving, single-step execution" closed-loop control mode, the system can continuously use the latest information to correct and adjust its control strategy, so as to effectively cope with prediction errors and various sudden disturbances, and ensure the optimality, robustness and safety of the control scheme in actual operation.

[0066] Further, the process of iteratively updating the energy efficiency model includes: at each client end, taking the dynamic reshaping coefficient as the true value label of supervised learning and taking the multi-dimensional feature vector as the input, performing multi-round training on the energy efficiency model by calculating the loss function between the predicted value and the true value label, and adopting the back propagation and gradient descent algorithm; after the local training is completed, the model weight update amount of the local copy relative to the global model before training is calculated and extracted, and the update amount is uploaded to the center server as the local model update parameter; after the center server collects the model weight update amounts from multiple clients, the federated average algorithm is adopted to perform weighted average on the multiple update amounts, and the aggregated global model update amount is calculated; the center server updates the maintained global optimization model using the aggregated global model update amount, and distributes the updated global optimization model to each client to replace the original local energy efficiency model, completes the iterative update, and the specific client-server interaction diagram is shown in Fig. 3

[0067] ​Specifically, the embodiment constructs a collaborative modeling framework that protects data privacy and continuously self-optimizes. On the edge computing device of each independent client (i.e., each central air conditioning system), the system performs local model training using the data generated in actual operation. Specifically, the system uses the generated multi-dimensional feature vector as the input of the model and uses the dynamic reshaping coefficient output by the energy efficiency model itself as the true value label of this training, which constitutes a self-supervised learning mechanism, enabling the model to continuously learn based on its own prediction and the dynamic response of the actual working condition. The training process uses the standard backpropagation algorithm and gradient descent optimizer (such as the Adam optimizer) in deep learning to adjust the weight parameters inside the energy efficiency model by calculating the difference between the reshaping coefficient predicted by the model and the reshaping coefficient as the label (for example, using the mean square error loss function). This local training process is iterated multiple times to ensure that the model learns the characteristics of the local data sufficiently. After local training is completed, the client does not upload the trained model itself, but calculates the change in all weight parameters of the local model relative to the global model downloaded from the server before training. The change in weight parameters is used as the local model update parameter, which is uploaded to the central server through an encrypted channel. The central server waits and collects model updates uploaded from multiple (e.g., more than a predetermined number) clients. Once enough updates are collected, the server performs a federated averaging algorithm, which weights and averages each update according to the amount of data used for local training by each client, thereby calculating an aggregated global model update that represents the learning achievements of all participants. Subsequently, the central server applies this aggregated update to the global optimization model it maintains to generate a new generation of global model with better performance and stronger generalization ability. Finally, the server distributes the updated global model to all clients, and the clients replace the original local energy efficiency model with the new model after receiving it.

[0068] To ensure the correct direction of the learning process, an external feedback signal based on actual operation results is introduced to guide the training. Specifically, the total power consumption and total refrigeration capacity measured in actual operation are used to calculate an energy efficiency ratio score, which is used to evaluate the dynamic reshaping coefficient output by the model in the last period. A policy gradient term is added to the loss function of the training model to adjust the model parameters to maximize this real energy efficiency ratio score. In this way, the iterative update of the model has an objective anchor point of pursuing higher energy efficiency, ensuring the final effect of federated learning.

[0069] The process of generating a global optimization model by the server aggregating parameters further comprises a dynamic weighted aggregation process based on energy efficiency feedback. In each aggregation cycle, the central server first collects model update amounts generated by local training of each client and average system energy efficiency ratio scores reflecting recent energy saving performance; the server sorts all clients in descending order according to the energy efficiency ratio scores, and assigns an initial, non-normalized contribution score to each client according to the sorting result (for example, the first ranked client gets the highest score, and the last ranked client gets the lowest score); the server performs normalization processing on the initial contribution scores of all clients to convert them into contribution weights, and uses the contribution weights to perform weighted summation on the model update amounts submitted by all clients to calculate a final global model update amount; and the weighted aggregated global update amount is applied to the global model maintained by the central server to complete iteration. By preferentially fusing model updates of clients with better actual energy saving effect, the convergence of the global model can be accelerated, and the interference of inefficient updates can be effectively suppressed, so that the global model generated finally has better performance and is more robust.

[0070] This local training-update-global aggregation-distributed model process constitutes a complete iteration cycle, which is continuously performed, so that the adaptability and prediction accuracy of the global model to various working conditions are continuously improved without contacting any original data of the clients.

[0071] By fusing physical mechanisms and real-time data, the accuracy of system state estimation is improved. Secondly, a two-stage model combining benchmark mode and dynamic remodeling is used to generate more reliable load prediction curves. Thirdly, based on the prediction, proactive optimal control is performed to replace the traditional passive response strategy. A federated learning framework is introduced to enable multiple clients to collaboratively train a shared global energy efficiency model while ensuring the privacy of system data. This is aimed at gathering the operation experience of multiple systems to build a control model that is more adaptable to diversified operating conditions and has more comprehensive decision basis, thereby improving the overall operation efficiency of the system.

[0072] Embodiment two:

[0073] To improve the operation efficiency of the central air conditioning system, a federated learning collaborative modeling method for optimizing the energy efficiency of the central air conditioning system is introduced.

[0074] Real-time power of four water chillers, running frequency of each water pump and cooling tower fan, supply and return water temperature of chilled and cooling water, and dry and wet bulb temperature collected by outdoor weather station. At the same time, multiple DN50 dynamic balance valves are installed in the air duct of the system terminal server cabinet, with a preset opening value of 4.5 turns. The balance valve debugging calculation program is built-in the edge computing node, which automatically calls functions, looks up tables and interpolates to calculate the KV value under the current setting as 30.9. In addition, through the personnel access record of the machine room management system and the CPU utilization of the server rack, the space occupancy state is quantified into three levels of high load, medium load and low load, and is coded with a one-hot code. All collected data are processed by Kalman filtering, normalization and timestamp alignment to form a standardized multi-dimensional feature vector.

[0075] The system measures the real-time pressure difference by the differential pressure sensor installed at both ends of the DN50 balance valve, combines the KV value calculated in the previous step, and inversely calculates the real-time flow through each terminal device by the fluid dynamics formula. The flow data of all terminals are combined and integrated, and the chilled water supply and return water temperature difference detected by the main pipeline (for example, supply water 7.2℃, return water 12.5℃), and the density and specific heat capacity of water calculated in real time according to the average water temperature, are used to accurately calculate the total real-time cooling load of the current data center as 2500kW.

[0076] The benchmark generation model uses the environmental data and IT load data accumulated in the past year to automatically identify and generate four typical daily load patterns through the LSTM encoder and K-means clustering algorithm (preset K=4), which are: summer IT business peak day mode, spring and autumn transition season day and night switching mode, winter low energy consumption stable operation mode and weekend and holiday low load mode.

[0077] The system detects that the outdoor wet bulb temperature is 28.5℃, CH101 and CH102 two water chillers are running at 85% load rate, the total system power is 700kW, and the space occupancy state is high load. This multi-dimensional feature vector containing real-time device data, environmental data and load data is input into the local energy efficiency model based on the Transformer architecture. The model analyzes the internal relationship between variables through its self-attention mechanism, and finally outputs a 16-dimensional dynamic reshaping coefficient vector.

[0078] The system selects the summer IT business peak day mode as the baseline. The parameter transformation network analyzes the dynamic reshaping coefficients generated in the previous step: among them, the time domain parameter indicates that due to an unplanned data migration task, the load decline trend from 18:00 to 20:00 in the evening is expected to be delayed by about one hour compared to the baseline mode; at the same time, the amplitude parameter indicates that the load in the next two hours will increase by 8% on the basis of the baseline mode to cope with sudden computing demand. By nonlinearly distorting the time axis of the baseline mode and multiplicatively modulating the local amplitude, the system generates a dynamic load prediction curve accurate to the next 24 hours.

[0079] Based on this high-precision load prediction curve, the system constructs a dynamic optimization problem with the goal of minimizing the total power consumption in the next 24 hours. In the current control period (15 minutes), the sequential quadratic programming algorithm is rolled out for solution. According to the principle of high-efficiency cooling system, the algorithm finds that although two cooling towers (2-to-2 host) are currently turned on, if one more cooling tower (3-to-2 host) is added and the overall fan speed of the three towers is appropriately reduced, the cooling water inlet temperature can be significantly reduced from 32°C to 29.8°C with very little increase in fan energy consumption. This will greatly improve the operating efficiency (COP) of the two water chillers; therefore, the system generates and outputs the optimal coordinated control instructions for the current period: start the cooling tower CT103, and set the fan frequency of CT101, CT102, and CT103 to 38Hz; at the same time, reduce the frequency of the chilled water pumps CDWP101 and CDWP102 from 45Hz to 42Hz.

[0080] After the day's operation is completed, the edge computing node of the data center (client A) uses the thousands of multi-dimensional feature vectors accumulated throughout the day as input and their corresponding dynamic reshaping coefficients (as labels) to train the local energy efficiency model for multiple rounds. After the training is completed, the model weight update is calculated and encrypted and uploaded to the central server located in the cloud. The server also receives model updates from two other data centers of client B and client C on the same day. The server performs federated averaging algorithm to combine the three model updates from different climate and load characteristics regions.

[0081] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system, characterized in that, include: Acquire system operating status data such as differential pressure data of balancing valves, preset KV value data, equipment operating data, supply and return water temperatures, and environmental data; The pressure difference of the balancing valve and the KV value are combined for calculation to obtain real-time flow data. The real-time load value is then calculated by combining the real-time flow data with the supply and return water temperatures. Environmental data is input into the baseline generation model to obtain the load pattern; Input equipment operation data, environmental data, real-time load values, and real-time flow data into the energy efficiency model to generate dynamic reshaping coefficients; By using a parameter transformation network, the dynamic reshaping coefficient is used as the transformation parameter of the network to perform time axis distortion and local amplitude modulation on the load mode, generating a load prediction curve. Based on the load prediction curve, the optimal control algorithm is used to solve the problem, with the goal of minimizing energy consumption, and a coordinated control command is generated to adjust the operating frequency of the variable frequency pump. On the client-side, the energy efficiency model is trained and updated parameters are generated using the dynamic reshaping coefficient as a label. After the server aggregates parameters to generate a global optimization model, it distributes the model to the client and iteratively updates the energy efficiency model.

2. The federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system according to claim 1, characterized in that, The acquisition of system operating status data, including balance valve differential pressure data, preset KV value data, equipment operating data, supply and return water temperatures, and environmental data, includes: the equipment operating data includes the operating frequency and real-time power of the variable frequency water pump; the environmental data includes indoor and outdoor temperature and humidity and space occupancy status; for the balance valve differential pressure data and the real-time power, a Kalman filter algorithm is used for noise reduction and smoothing; for the space occupancy status, one-heat encoding is used to convert it into numerical features; the preset KV value data, the operating frequency, the supply and return water temperatures, and the indoor and outdoor temperature and humidity are normalized to eliminate the influence of different physical dimensions; the preprocessed data stream is synchronized and aligned based on a unified timestamp, and time series interpolation is used to fill in missing values ​​in the sequence.

3. The federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system according to claim 1, characterized in that, The process of calculating the real-time load value includes: substituting the differential pressure data of the balancing valve and the preset KV value into the valve flow coefficient calculation formula to calculate the theoretical flow reference value; dynamically adjusting the internal correction parameters according to the real-time deviation between the theoretical and actual flow through the online adaptive corrector, and outputting the real-time correction value; and during real-time calculation, weighted summing of the theoretical flow reference value and the real-time correction value to generate the real-time flow data. The average temperature of the current fluid medium is calculated based on the supply and return water temperature data; a reference state vector is constructed by predefined discrete temperature nodes and corresponding standard fluid density and specific heat capacity parameters, and an interpolation weight vector is generated based on the position of the average temperature among the discrete temperature nodes; the fluid density and specific heat capacity corresponding to the current average temperature are calculated by performing a vector dot product operation between the interpolation weight vector and the reference state vector; the real-time load value is calculated and output by combining the real-time flow data, fluid density and specific heat capacity and supply and return water temperature difference through the thermodynamic energy balance equation.

4. The federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system according to claim 1, characterized in that, The process of inputting environmental data into a benchmark generation model to obtain load patterns includes: dividing the environmental data into periodic segments to form environmental data fragments as independent samples; inputting each environmental data fragment into the benchmark generation model, which is an encoder composed of a long short-term memory network, to compress and extract the environmental data fragments into low-dimensional latent vectors; aggregating the low-dimensional latent vectors and performing cluster analysis using the K-means clustering algorithm in the low-dimensional latent vector space to identify statistically significant typical daily patterns; and outputting the centroid vector of each cluster as the load pattern.

5. The federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system according to claim 1, characterized in that, The process of generating dynamic reshaping coefficients includes: aligning the device operation data, environmental data, real-time load values, and real-time traffic data on the time axis, and at each synchronized time sampling point, concatenating the values ​​from all data sources in a preset order to construct a multi-dimensional feature vector; inputting the multi-dimensional feature vector into an energy efficiency model based on the Transformer architecture, and performing weighted summation and information fusion on the multi-dimensional feature vector; the energy efficiency model maps and outputs the fused high-dimensional features through a fully connected output layer to generate dynamic reshaping coefficients.

6. The federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system according to claim 1, characterized in that, The process of generating the load prediction curve includes: extracting elements corresponding to a first preset index range from the dynamic reshaping coefficients as time-domain parameters controlling the nonlinear transformation of the time axis, and extracting elements corresponding to a second preset index range as amplitude parameters controlling local amplitude changes; inputting the time-domain parameters into the time-domain transformation module of the parameter transformation network to generate a one-dimensional sampling grid, and using a differentiable interpolation algorithm to resample the load pattern according to the one-dimensional sampling grid to generate an intermediate load pattern with a nonlinearly distorted time axis; inputting the amplitude parameters into the amplitude modulation module of the parameter transformation network to generate a multiplicative mask of the same length as the intermediate load pattern and containing a local scaling factor; and multiplying the multiplicative mask point-by-point with the intermediate load pattern to obtain the load prediction curve.

7. The federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system according to claim 1, characterized in that, The process of generating coordinated control commands includes: using the load prediction curve as a system demand constraint, and taking the minimization of the energy consumption of the variable frequency pump as the optimization objective, constructing a multivariable, constrained dynamic optimization problem; within each control cycle, using a sequential quadratic programming algorithm to solve the dynamic optimization problem in a rolling manner to obtain the optimal control sequence; and extracting the first control command from the optimal control sequence as the coordinated control command for the current moment for output.

8. The federated learning collaborative modeling method for energy efficiency optimization of a central air conditioning system according to claim 1, characterized in that, The iterative update process of the energy efficiency model includes: on each client, the dynamic reshaping coefficient is used as the ground truth label for supervised learning, and the multidimensional feature vector is used as input. The loss function between the predicted value and the ground truth label is calculated, and the energy efficiency model is trained multiple times using backpropagation and gradient descent algorithms. After local training is completed, the model weight update amount of the local copy relative to the global model before training is calculated and extracted, and the update amount is uploaded to the central server as the local model update parameter. After collecting the model weight update amounts from multiple clients, the central server uses a federated averaging algorithm to perform a weighted average of the multiple update amounts to calculate the aggregated global model update amount. The central server uses the aggregated global model update amount to update the maintained global optimization model, and distributes the updated global optimization model to each client to replace the original local energy efficiency model, thus completing the iterative update.

Images