An air conditioner load cluster prediction method based on a double-layer soft gate hybrid Gaussian process

Abstract

The application relates to the field of power load prediction, and provides an air conditioner load cluster prediction method based on a double-layer soft-gated hybrid Gaussian process: multi-dimensional input features for prediction are constructed, and the data is preprocessed; an unobservable device start-stop state is mapped into observable aggregated power dynamic change; an upper-layer daily potential mode division model is established, and a lower-layer soft-gated multi-expert Gaussian process model is established; a neural network mean function and an additive combination kernel function are constructed to jointly represent nonlinear trends, time smoothness, local periodicity, temperature correlation and random disturbance in load change; a Monte Carlo expectation maximization algorithm is used to learn model parameters, and an air conditioner load cluster short-term prediction result and a prediction uncertainty interval are output based on the trained model. The application improves the short-term prediction precision and generalization ability of the air conditioner load cluster.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting, and in particular to a method for forecasting air conditioning load clusters based on a two-layer soft-gated mixed Gaussian process. Background Technology

[0002] Air conditioning load clusters represent large-scale demand-side resources with significant adjustment potential in energy-consuming scenarios such as buildings, industrial parks, and schools. As the installed capacity and power generation share of renewable energy sources continue to increase, the uncertainty and volatility of power system operation are growing, placing higher demands on the state perception, adjustability potential assessment, and short-term power forecasting of flexible load-side resources. Accurately predicting load changes in air conditioning load clusters in the short term provides a foundation for demand response, operational boundary quantification, and refined power system scheduling.

[0003] Existing methods for predicting air conditioning or building loads mainly include autoregressive integral moving average models, support vector machines, random forests, long short-term memory networks, gated recurrent units, bidirectional long short-term memory networks, and improved models based on clustering or deep learning. While these methods have some applicability in conventional load prediction scenarios, they still have shortcomings in air conditioning load cluster scenarios. On the one hand, air conditioning clusters are affected by factors such as equipment start-up and shutdown, user behavior, ambient temperature, ambient humidity, weekday attributes, and holidays, resulting in load curves exhibiting multi-modal, nonlinear, random, and non-stationary characteristics. On the other hand, a single model cannot simultaneously characterize the heterogeneous response mechanisms under different operating conditions, leading to insufficient prediction accuracy for peak intervals, rapidly changing intervals, and adjustable potential related states.

[0004] While some neural network models can improve point prediction accuracy, their ability to interpret prediction uncertainties, temporal periodic structures, and physical influencing factors is limited. Some Gaussian process models can provide descriptions of uncertainties, but using a single Gaussian process, a fixed mean function, or a simple kernel function makes it difficult to adapt to nonlinear coupling relationships and multi-mode operating characteristics under multi-dimensional input conditions. Therefore, it is necessary to provide a short-term prediction method for air conditioning load clusters that combines daily operating mode division, power increment modeling caused by equipment start-up and shutdown, soft-gated multi-expert collaborative modeling, neural network mean functions, additive combined kernel functions, and uncertain outputs, relying solely on aggregated power and external driving information. Summary of the Invention

[0005] Based on the above, the purpose of this invention is to provide a method for predicting air conditioning load clusters based on a two-layer soft-gated mixed Gaussian process. Under the premise of obtaining the aggregated power of the air conditioning cluster and external driving information, the method jointly models the daily operating mode of the air conditioning cluster load, the power increment caused by equipment start-up and shutdown, the heterogeneous response within the mode, and the prediction uncertainty, thereby improving the accuracy, robustness, and interpretability of short-term prediction of air conditioning load clusters and providing support for demand response and refined scheduling.

[0006] To achieve the above objectives, this invention provides a method for predicting clustered air conditioning loads based on a two-layer soft-gated Gaussian mixture process, the method comprising:

[0007] Step S1: Obtain air conditioning load cluster operation data and external driving data, construct multidimensional input features for prediction, and preprocess the data.

[0008] Specifically, it includes:

[0009] Step S11: Obtain aggregated power data of the air conditioning load cluster through a non-intrusive load monitoring terminal or energy consumption acquisition system, and simultaneously obtain external driving information such as ambient temperature, humidity, time and location, weekday attribute, weekend identifier and holiday identifier.

[0010] Step S12: Perform missing value processing, outlier processing, time alignment, sample construction, and standardization on the collected data to form a multidimensional input sample containing historical load, meteorological characteristics, and calendar characteristics.

[0011] Step S2: Construct a power increment observation model based on the start-up and shutdown mechanism of air conditioning equipment, and map the start-up and shutdown state of the equipment, which cannot be directly observed, into the dynamic change of aggregate power that can be observed.

[0012] Specifically, it includes:

[0013] Step S21: The total power of the air conditioning cluster is expressed as the sum of the power of individual devices in operation, and the number of operating units and the changes in device start-up and shutdown are used as implicit variables reflecting the group's operating status.

[0014] Step S22: Use the difference in aggregated power between adjacent time points as the observation variable, and construct linear coefficients to characterize the remaining operable capacity and the scale of the operating load, respectively.

[0015] Step S23: Based on the enabled and disabled implicit functions, describe the start-up and shutdown ratio of the equipment, so that the model can characterize the dynamic response of the air conditioning cluster by aggregating power increments without directly observing the status of individual equipment.

[0016] Step S3: Establish a higher-level daily potential pattern partitioning model to identify the operating mode of the air conditioning load curves on different dates, so as to reduce the impact of sample heterogeneity on the modeling results.

[0017] Specifically, it includes:

[0018] Step S31: Using the daily load curve as a sample, introduce daily potential pattern variables to classify the overall operating conditions on different dates.

[0019] Step S32: Set potential mode probabilities for the on-side and off-side respectively, so that the load curves of different dates can be classified into the corresponding typical operating modes, thereby reducing curve confounding and sample heterogeneity.

[0020] Step S4: Establish a lower-level soft-gated multi-expert Gaussian process model. Under a given operating mode, adaptively fuse multiple expert Gaussian processes based on input-related gating weights to characterize the heterogeneous start-stop response within the mode.

[0021] Specifically, it includes:

[0022] Step S42: Construct the input-related gating function using a single hidden layer neural network, and obtain the gating weights of each expert model through normalization processing.

[0023] Step S43: Perform weighted fusion of multiple expert Gaussian processes according to the gating weights, so that the model can continuously adjust the contributions of different experts as time, temperature and calendar attributes change.

[0024] Step S5: Construct a neural network mean function and an additive combined kernel function to jointly characterize the nonlinear trend, time smoothness, local periodicity, temperature correlation and random disturbance in load changes.

[0025] Specifically, it includes:

[0026] Step S51: Use a neural network to construct the Gaussian process mean function instead of the traditional basis function expansion method, so as to enhance the model's ability to express multidimensional input variables and complex nonlinear coupling relationships.

[0027] Step S52: Construct an additive combined kernel function consisting of a time smoothing term, a temperature-dependent term, a local periodic term, and a white noise term to characterize the smoothness, periodicity, temperature influence, and random disturbances of the load sequence.

[0028] Step S53: Under the action of soft gating weights, a combined mean and combined covariance structure is formed to achieve a joint characterization of load trend changes and fluctuation characteristics under different operating conditions.

[0029] Step S6: The Monte Carlo expectation-maximization algorithm is used to learn the model parameters, and the short-term prediction results and prediction uncertainty interval of the air conditioning load cluster are output based on the trained model.

[0030] Specifically, it includes:

[0031] Step S61: Construct the observation likelihood based on the power increment observation model, and combine the upper-level latent mode variables and the lower-level implicit function priors to form a complete joint probability model.

[0032] Step S62: In the expected step, latent pattern variables are sampled and posterior samples are performed on the Gaussian process implicit function to obtain posterior samples for parameter updates.

[0033] Step S63: In the maximization step, update the mixture probability, gated network parameters, mean network parameters, kernel function hyperparameters, and noise parameters based on the sampling results to obtain the trained prediction model.

[0034] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0035] The short-term forecasting method for air conditioning load clusters provided by this invention takes the power changes caused by the start-up and shutdown of air conditioning equipment as the observation object, and can characterize the dynamic response of the cluster under the condition of relying only on aggregated power and external driving information. It reduces the impact of mixed operating conditions on modeling by upper-level daily potential mode partitioning; it adaptively describes the heterogeneous response within the mode through lower-level soft-gated multi-expert Gaussian processes; and it jointly characterizes nonlinear trends, time-series periodic patterns, and prediction uncertainties through neural network mean functions and additive combined kernel functions. This method can improve the accuracy and generalization ability of short-term forecasting of air conditioning load clusters, and provide a reliable basis for air conditioning cluster demand response, flexible resource assessment, and refined power system scheduling. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart of the short-term forecasting method for air conditioning load clusters according to an embodiment of the present invention;

[0038] Figure 2 This is a structural diagram of a two-layer soft-gated Gaussian mixture process model according to an embodiment of the present invention;

[0039] Figure 3This is a diagram showing the short-term forecast results and uncertainty interval of the air conditioning load cluster according to an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The purpose of this invention is to provide a method for predicting air conditioning load clusters based on a two-layer soft-gated hybrid Gaussian process. Under the premise of obtaining the aggregated power and external driving information of the air conditioning cluster, the method performs daily-level operation mode division, power increment observation modeling, soft-gated multi-expert Gaussian process modeling, and parameter learning to achieve short-term power prediction of the air conditioning load cluster.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] like Figure 1 As shown, this invention discloses a method for predicting clustered air conditioning loads based on a two-layer soft-gated Gaussian mixture process, the method comprising:

[0044] Step S1: Obtain air conditioning load cluster operation data and external driving data, construct multidimensional input features for prediction, and preprocess the data.

[0045] Step S11: In air conditioning cluster application scenarios such as buildings and parks, collect the aggregated power of the air conditioning cluster through a non-intrusive load monitoring terminal or energy management system, and at the same time collect external driving information related to air conditioning operation such as ambient temperature, humidity, daytime, weekday attribute, weekend label, and holiday label.

[0046] Step S12: Perform time alignment, missing value imputation, outlier removal, and standardization on the collected data, and construct a daily sample curve according to the prediction time scale. For any discrete time t, construct a multidimensional input feature vector X. t .

[0047] Step S2: Construct a power increment observation model based on the start-up and shutdown mechanism of air conditioning equipment, and map the start-up and shutdown state of the equipment, which cannot be directly observed, into the dynamic change of aggregate power that can be observed.

[0048] Step S21: Consider an air conditioning cluster consisting of N air conditioning units. Let the on / off state of the i-th air conditioner at time t be s. i,t , Its operating power is p i,t Then the cluster aggregation power can be expressed as: In the formula, s i,t The value is used to characterize whether the i-th air conditioner is in operation. Since the start-up and shutdown status of individual devices is usually not directly observable in real-world scenarios, this invention treats the number of operating units and the change in start-up and shutdown status as implicit states, and reflects the evolution of the group state through changes in aggregated power.

[0049] Step S22: Based on the characteristic that the current power of the air conditioning cluster is determined by the power of the previous moment, the power increase caused by equipment startup, the power decrease caused by equipment shutdown, and random disturbances, the cluster power recursion process is expressed as follows: ;in, and These represent the power changes caused by the device being turned on and off at time t, respectively. This represents the equivalent disturbance caused by the combined effects of single-machine power fluctuations and unmodeled factors.

[0050] Step S23: Further convert the state changes of the air conditioning cluster into incremental observations in the power domain, defining the power increment and linear coefficient as follows: ;

[0051] The short-time power evolution process of the air conditioning cluster can be expressed as: ;in, and These represent enabling and disabling implicit functions, respectively. This is a mapping function used to map the implicit function output to an on or off ratio. Therefore, the dynamic response of the air conditioning cluster can be characterized by aggregated power increments without directly observing the internal state of the equipment.

[0052] Step S3: Establish a higher-level daily potential pattern partitioning model to identify the operating mode of the air conditioning load curves on different dates, so as to reduce the impact of sample heterogeneity on the modeling results.

[0053] Step S31: Treat each day's air conditioning load curve as a daily sample curve, and denote the multidimensional input sequence corresponding to the m-th daily sample curve as x. m Because operating conditions, user usage patterns, and environmental conditions vary across different dates, this invention introduces daily-level latent pattern variables. The sample curves are divided into modes, where s represents the on or off side.

[0054] Step S32: Set the daily-level latent pattern variable to follow a categorical distribution, and its prior can be expressed as: ; ;in, This represents the mixed probability vector for the corresponding side. Through the upper-level latent pattern variables, the model can divide the load curves of different dates into several typical operating modes, thereby reducing the impact of mixed operating conditions on subsequent forecasting and modeling.

[0055] Step S4: Establish a lower-level soft-gated multi-expert Gaussian process model, such as... Figure 2 As shown, under a given operating mode, multiple expert Gaussian processes are adaptively fused based on input-related gating weights to characterize the heterogeneous start-stop responses within the mode.

[0056] Step S41: Given the m-th daily-level sample curve, the upper-level latent mode k, and the side s, represent the implicit function of the open or closed side as a soft combination of multiple expert Gaussian processes, specifically: Where G is the number of experts in each mode. Let g be the gating weight of the g-th expert in the k-th mode, and satisfy the following conditions: , This corresponds to the expert implicit function.

[0057] Step S42: Model each expert implicit function as a Gaussian process to characterize the trend and uncertainty under the local response mechanism: ;in, For expert mean function, For expert kernel functions.

[0058] Step S43: To ensure that expert contributions can be continuously adjusted according to changes in input conditions, a single hidden layer neural network is used to construct a gating function, and the gating weights are obtained through softmax normalization. , .

[0059] Through the aforementioned soft gating mechanism, the model can dynamically call more suitable local experts based on input information such as time, temperature, and calendar attributes, avoiding boundary discontinuity problems caused by hard switching and enhancing the ability to characterize the heterogeneous start-up and shutdown patterns of air conditioning clusters.

[0060] Step S5: Construct a neural network mean function and an additive combined kernel function to jointly characterize the nonlinear trend, time smoothness, local periodicity, temperature correlation and random disturbance in load changes.

[0061] Step S51: For each expert Gaussian process, a neural network is used to construct the mean function, replacing the traditional linear combination of basis functions. The mean function of the g-th expert can be expressed as: , .

[0062] This mean function can directly learn the nonlinear trend of load changes from multidimensional input features, and is suitable for air conditioning cluster load forecasting under the combined effects of environmental factors, time factors, and user behavior factors.

[0063] Step S52: At the covariance level, construct an additive combined kernel function consisting of a time smoothing term, a temperature-dependent term, a local periodic term, and a white noise term. Let time difference be defined for any expert g. temperature difference Then the kernel function can be written as: In the formula, the first term is used to characterize long-term time smoothness, the second term is used to characterize the smoothness correlation in the temperature dimension, the third term is used to characterize local periodicity, and the fourth term is used to absorb random disturbances.

[0064] Step S53: Under the condition that the experts are independent of each other, the combined implicit function after soft-gated weighting can still form the combined mean and combined covariance structure; for the implicit function of the m-th sample curve on the discrete-time grid, it can be expressed as: ;in, and These are the combined mean vector and combined covariance matrix, calculated based on soft-gated weights, respectively. Using this structure, the present invention can adaptively model the nonlinear trends, periodic characteristics, and uncertainties within a daily-level model, based on the model's partitioning.

[0065] Step S6: The Monte Carlo expectation-maximization algorithm is used to learn the model parameters, and the short-term prediction results and prediction uncertainty interval of the air conditioning load cluster are output based on the trained model.

[0066] Step S61: Based on the power increment observation equation, given the implicit functions on the turn-on and turn-off sides, the observation likelihood of the m-th sample curve can be expressed as: .

[0067] By further combining the upper-level latent pattern variables, the lower-level implicit function priors, and the model parameters, a complete joint probability model can be formed. Since the implicit function enters the observation mean term through a mapping function, the posterior distribution no longer has an analytical form; therefore, the Monte Carlo expectation-maximization algorithm is used for parameter learning.

[0068] Step S62: Let the model parameters be uniformly denoted as... The objective function of the Monte Carlo expectation-maximization algorithm can be expressed as: ;

[0069] Since the conditional expectation is difficult to calculate analytically, S sets of samples are obtained using posterior sampling, and the Monte Carlo approximation is applied: .

[0070] In the expected step, latent pattern variables are sampled and implicit functions are posteriorly sampled to obtain a sample set for parameter updates.

[0071] Step S63: In the maximization step, the responsibility of the sample curves for different latent modes is calculated based on the sampling results, and the mixture probability is updated. The mixture probability can be updated as follows: .in, This represents the posterior responsibility of the m-th sample curve belonging to the k-th latent mode. For the gated network parameters, mean network parameters, kernel function hyperparameters, and noise parameters, a differentiable optimization objective can be constructed based on the sampling results, and a gradient optimization method can be used for joint updates.

[0072] Step S64: After model training is complete, the input features at the prediction time are substituted into the upper-layer latent mode model and the lower-layer soft-gated multi-expert Gaussian process model to obtain the corresponding on-side and off-side implicit function prediction results. Based on the power increment observation model, the predicted power increment and future aggregated power are obtained. , In this context, variables marked with superscripts represent predicted values. By utilizing the prediction covariance of a Gaussian process, prediction confidence intervals can be further output, thus providing an uncertainty reference for demand response execution, assessment of air conditioning load regulation potential, and refined power system scheduling.

[0073] To further illustrate the predictive effect of the method of the present invention, the actual operating data of the air conditioning load cluster of a certain school was selected as the dataset of the example. The historical aggregated power, ambient temperature, humidity, intraday time, weekday attribute, weekend identifier and holiday identifier were used as input features to make short-term predictions on the aggregated power of the air conditioning load cluster in the future period. Figure 3 The prediction results and uncertainty intervals of the method of the present invention are given, where the horizontal axis is the prediction time, the vertical axis is the aggregated power of the air conditioning load cluster, the true value curve is used to represent the actual collected aggregated power, the predicted value curve is used to represent the short-term predicted power output by the method of the present invention, and the shaded area or upper and lower boundary curves represent the 95% prediction confidence interval obtained based on the prediction covariance of the Gaussian process.

[0074] Depend on Figure 3 As can be seen, the prediction curve output by the method of the present invention can follow the changes of the actual load curve well, and can maintain high consistency in the load rising interval, peak interval and falling interval. At the same time, the prediction confidence interval can cover the main load fluctuation range, indicating that the method of the present invention can not only output the short-term point prediction results of the air conditioning load cluster, but also give the prediction uncertainty range, thereby providing a reference for demand response execution, adjustable potential assessment and refined power system scheduling.

Claims

1. A method for predicting clustered air conditioning loads based on a two-layer soft-gated Gaussian mixture process, characterized in that, The method includes: Step S1: Obtain air conditioning load cluster operation data and external driving data, construct multidimensional input features for prediction, and preprocess the data; Step S2: Construct a power increment observation model based on the start-up and shutdown mechanism of air conditioning equipment, and map the start-up and shutdown state of the equipment, which cannot be directly observed, into the dynamic change of aggregate power that can be observed. Step S3: Establish a higher-level daily potential pattern partitioning model to identify the operating mode of the air conditioning load curves on different dates, so as to reduce the impact of sample heterogeneity on the modeling results; Step S4: Establish a lower-level soft-gated multi-expert Gaussian process model. Under a given operating mode, adaptively fuse multiple expert Gaussian processes based on input-related gating weights to characterize the heterogeneous start-stop response within the mode. Step S5: Construct a neural network mean function and an additive combined kernel function to jointly characterize the nonlinear trend, time smoothness, local periodicity, temperature correlation and random disturbance in load changes; Step S6: The Monte Carlo expectation-maximization algorithm is used to learn the model parameters, and the short-term prediction results and prediction uncertainty interval of the air conditioning load cluster are output based on the trained model.

2. The air conditioning load cluster prediction method based on a two-layer soft-gated mixed Gaussian process according to claim 1, characterized in that, Step S1 specifically includes: Step S11: Obtain aggregated power data of the air conditioning load cluster, and simultaneously obtain ambient temperature, humidity, time and location, weekday attribute, weekend identifier and holiday identifier; Step S12: Perform missing value processing, outlier processing, time alignment, sample construction, and standardization on the collected data to form an input sample containing historical load, meteorological characteristics, and calendar characteristics.

3. The air conditioning load cluster prediction method based on a two-layer soft-gated mixed Gaussian process according to claim 1, characterized in that, Step S2 specifically includes: Step S21: The total power of the air conditioning cluster is expressed as the sum of the power of individual devices in operation, and the number of operating units and the change in device start-up and shutdown are used as implicit variables reflecting the group's operating status. Step S22: Use the difference in aggregated power between adjacent time points as the observed variable, and construct linear coefficients to characterize the remaining operable capacity and the scale of the operating load; Step S23: Use the enabled and disabled implicit functions to describe the device start-up and shutdown ratios, so as to characterize the dynamic response of the air conditioning cluster by aggregating power increments.

4. The air conditioning load cluster prediction method based on a two-layer soft-gated mixed Gaussian process according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Using the daily load curve as a sample, introduce daily potential pattern variables to classify the overall operating conditions under different dates. Step S32: Set potential mode probabilities for the start-up side and the shutdown side respectively, so that the load curves of different dates belong to the corresponding typical operating modes.

5. The air conditioning load cluster prediction method based on a two-layer soft-gated mixed Gaussian process according to claim 1, characterized in that, Step S4 specifically includes: Step S41: Set up multiple expert Gaussian processes within each potential operating mode, each expert Gaussian process being used to describe a local start-stop response mechanism; Step S42: Construct the input-related gating function using a single hidden layer neural network, and obtain the gating weights of each expert model through normalization processing; Step S43: Perform weighted fusion of multiple expert Gaussian processes according to the gating weights, so that the model continuously adjusts the contributions of different experts as time, temperature and calendar attributes change.

6. The air conditioning load cluster prediction method based on a two-layer soft-gated mixed Gaussian process according to claim 1, characterized in that, Step S5 specifically includes: Step S51: Construct a Gaussian process mean function using a neural network to enhance the model's ability to express multidimensional input variables and complex nonlinear coupling relationships; Step S52: Construct an additive combined kernel function consisting of a time smoothing term, a temperature-dependent term, a local periodic term, and a white noise term to characterize the smoothness, periodicity, temperature influence, and random disturbances of the load sequence; Step S53: Under the action of soft gating weights, a combined mean and combined covariance structure is formed to achieve a joint characterization of load trend changes and fluctuation characteristics under different operating conditions.

7. The air conditioning load cluster prediction method based on a two-layer soft-gated Gaussian mixture process according to claim 1, characterized in that, Step S6 specifically includes: Step S61: Construct the observation likelihood based on the power increment observation model, and combine the upper-level latent mode variables and the lower-level implicit function priors to form a complete joint probability model; Step S62: In the expected step, latent pattern variables are sampled and posterior samples are performed on the Gaussian process implicit function to obtain posterior samples for parameter updates. Step S63: In the maximization step, update the mixture probability, gated network parameters, mean network parameters, kernel function hyperparameters, and noise parameters based on the sampling results to obtain the trained prediction model.

8. The air conditioning load cluster prediction method based on a two-layer soft-gated Gaussian mixture process according to claim 1, characterized in that, The output short-term prediction results and prediction uncertainty intervals for the air conditioning load clusters specifically include: substituting the multidimensional input features at the prediction time into the trained model to obtain the implicit function prediction results for the on-side and off-side, calculating the predicted power increment and recursively obtaining the future aggregated power, and generating the prediction confidence interval based on the Gaussian process prediction covariance.

Images