A feedback control system for optimizing the performance of grid-interactive building systems using a time-series-based model.

Abstract

This provides a feedback control system for optimizing the performance of a grid-interactive building (GIB) system configured to regulate the indoor environment of a building. [Solution] The feedback control system comprises a time-series underlying model that predicts disturbances affecting the building's energy consumption, and a probabilistic feedback controller configured to determine control inputs to the GIB system by evaluating multiple control actions for the GIB system based on the predicted disturbances, wherein the probabilistic feedback controller maximizes the likelihood that the control inputs achieve desired indoor environmental conditions while minimizing energy consumption. The feedback mechanism of the feedback control system fine-tunes the time-series underlying model and updates the probabilistic feedback controller using real-time data including indoor environmental conditions, actual energy consumption, and external factors.

Description

Technical Field

[0001] The present disclosure generally relates to control systems, and more specifically to feedback control systems and methods for optimizing the performance of grid-interactive building (GIB) systems using time-series based models.

Background Art

[0002] Optimization-based control and estimation techniques such as model predictive control (MPC) enable a model-based design framework that can directly take into account system dynamics and constraints. MPC is used to control dynamic systems of various complexities in many applications, and these systems are described by a set of non-linear differential equations, i.e., a system of ordinary differential equations (ODEs), differential-algebraic equations (DAEs), or partial differential equations (PDEs). Examples of such systems include production lines, automotive engines, robots, numerical control machining, satellites, and generators.

[0003] MPC is based on real-time finite horizon optimization of a model of the system. MPC has the function of predicting future events and taking appropriate control actions. This is achieved by optimizing the operation of the system over a finite future time horizon with constraints and implementing control only over the current time step.

[0004] MPC can predict changes in state variables of a modeled system caused by changes in control variables. State variables define the state of the system; that is, the state of the controlled system is the smallest set of state variables in the state-space representation of the control system that can represent the overall state of the system at a given time. For example, if the controlled system is an autonomous vehicle, the state variables may include the vehicle's position, velocity, and direction of travel. Using the system model, current system measurements and / or state estimates, the vehicle's current state, and state and control constraints, MPC calculates future changes in the vehicle's state. These changes are calculated to maintain the state near a target that is constrained on both the control variables and state variables. Typically, MPC outputs only the first change of each control variable that will be realized by the actuators of the controlled system, and repeats the calculation when the next change is needed.

[0005] Many systems under control are partially unknown, or at least uncertain. For example, when controlling a vehicle, the maximum friction between the tires and the road is not precisely known, nor is the dependence of friction on the vehicle's state, such as its speed. Typically, such uncertainties are estimated concurrently with the MPC to give the MPC better knowledge about the model it controls. While the MPC exhibits inherent robustness through feedback, such a controller does not directly account for uncertainty, and therefore cannot guarantee the satisfaction of safety-critical constraints in the presence of model uncertainties or external disturbances. One alternative is a robust MPC, which relies on optimizing the control policy under worst-case scenarios where bounded-range uncertainty exists. However, because the probability of worst-case scenarios occurring is extremely low, a robust MPC may lead to conservative control performance.

[0006] Another type of MPC is the stochastic model predictive control (SMPC), in which the system uncertainty is modeled to have a distribution, for example, a Gaussian distribution with a mean (center) and covariance (uncertainty). SMPCs are intended to reduce the conservatism of robust MPCs by directly incorporating the probabilistic description of uncertainty into the formulation of the optimal control problem (OCP). In some implementations, SMPCs require that the constraints be satisfied with a certain probability, i.e., by formulating so-called opportunity constraints that allow for a specified but non-zero probability of constraint violation. In addition, SMPCs are advantageous in settings where high performance of closed-loop operation is achieved near the boundary of the plant's feasible domain. In general cases, opportunity constraints are computationally difficult and typically require approximate formulations.

[0007] In addition to many systems with uncertain parameters or components, the disturbances acting on them are often also uncertain. While there are several methods for estimating the uncertainty of system parameters or components, these methods depend on the system's dynamics, whereas the uncertainty of disturbances may be independent of system dynamics.

[0008] In addition, some SMPC solvers assume that uncertainty is predetermined offline, i.e., before the controller is executed. In the case of disturbances, such an assumption is overly restrictive, because in many applications, uncertainty changes over time and therefore cannot be predetermined offline before the SMPC is executed.

[0009] Therefore, control systems are still needed to manage systems that involve uncertainty in the disturbances acting on them. [Overview of the project]

[0010] The objective of some embodiments is to provide systems and methods for controlling and optimizing the performance of a system under the influence of the uncertainty of disturbances acting on the system. In one embodiment, the system is a grid-interactive building (GIB) system. The GIB system includes one or more of the following: a heating, ventilation, and air conditioning (HVAC) system, a renewable energy source, and a power grid. The HVAC system, the renewable energy source, and the power grid are electrically interconnected.

[0011] HVAC systems are installed within a building and configured to regulate the building's indoor environment. For example, an HVAC system is configured to maintain the indoor temperature and / or humidity within a desired range or at a desired value. The indoor environment of a building refers to the space of the building's rooms or floors. In one embodiment, the HVAC system is powered by electrical energy from the power grid. In addition, the building is supplied with electrical energy from the power grid. The power grid is a network of transmission and distribution systems that send electricity from power plants to homes, businesses, and other users. It is the infrastructure that enables the generation, transmission, and distribution of electrical energy over long distances.

[0012] The objective of some embodiments is to optimize the performance of the GIB system. For example, the objective of some embodiments is to minimize the electrical energy supplied from the power grid by utilizing renewable energy sources. Renewable energy sources include on-site renewable energy sources and on-site energy storage systems, such as photovoltaic systems and batteries associated with the building. Also, the objective of some embodiments is to minimize the energy consumption of the HVAC system and the building. In addition, the objective of some embodiments is to minimize the energy consumption of the HVAC system while maintaining the indoor environment temperature and humidity at desired values.

[0013] To achieve such objectives, embodiments of the present disclosure provide a feedback control system. The feedback control system is communicatively coupled to a GIB system. In some other embodiments, the feedback control system is integrated into the GIB system. The feedback control system includes a processor, memory, a time-series underlying model, a probabilistic feedback controller, and a feedback mechanism. The processor may be a single-core processor, a multi-core processor, a computing cluster, or any number of other configurations. The memory may include random-access memory (RAM), read-only memory (ROM), flash memory, or any other suitable memory system. In addition, in some embodiments, the memory may be implemented using a hard drive, an optical drive, a thumb drive, an array of drives, or any combination thereof. In some embodiments, the time-series underlying model, the probabilistic feedback controller, and the feedback mechanism are modules of the feedback control system and are executed by the processor.

[0014] A time-series-based model is a type of machine learning model specifically designed to handle, understand, and make predictions from time-series data. Time-series data typically consists of a sequence of data points ordered relative to time and is used to model phenomena that evolve over time. In this disclosure, the time-series-based model is a pre-trained model configured to predict disturbances affecting a building's energy consumption. Disturbances affecting a building's energy consumption include external factors describing weather conditions, occupancy patterns in the indoor environment, and external temperature fluctuations. The predicted disturbances are fed into a probabilistic feedback controller.

[0015] A probabilistic feedback controller is configured to determine control inputs to a GIB system by evaluating multiple control actions for the GIB system based on disturbances affecting the predicted energy consumption of the building. The control inputs are determined to maximize the likelihood of achieving desired indoor environmental conditions while minimizing energy consumption. Desired indoor environmental conditions include desired temperature and humidity for the indoor environment. Control inputs include, for example, the compressor speed of the HVAC system, the position of the expansion valve, and the speeds of the indoor and outdoor fans of the HVAC system. The probabilistic feedback controller is further configured to generate control commands for the actuators of the HVAC system based on the control inputs and to control the actuators of the HVAC system according to the control commands to operate the HVAC system according to the control inputs.

[0016] A probabilistic feedback controller controls the HVAC system according to control inputs determined by considering predicted uncertainties, thereby precisely achieving the desired temperature and humidity of the indoor environment while minimizing energy consumption for both the HVAC system and the building. This optimizes the performance of the GIB system.

[0017] Some embodiments are based on the understanding that the predictive capabilities of time-series-based models can be improved by fine-tuning them based on real-time data. To this end, some embodiments provide a feedback mechanism configured to fine-tune the time-series-based model based on real-time data.

[0018] To this end, the feedback mechanism is configured to collect real-time data. This real-time data includes indoor environmental conditions, actual energy consumption by the HVAC system, and external factors. Indoor environmental conditions include temperature, humidity, and carbon dioxide (CO2) levels in the indoor environment. External factors include ambient temperature, ambient humidity, and solar radiation. In addition, in some embodiments, the real-time data includes inputs targeting occupants, such as equipment energy consumption.

[0019] Furthermore, the feedback mechanism is configured to fine-tune the time-series underlying model based on real-time data. By fine-tuning the time-series underlying model based on real-time data, the accuracy of disturbance predictions is improved. In some embodiments, the feedback mechanism periodically collects real-time data and periodically updates the time-series underlying model. In this way, the feedback control system dynamically optimizes the performance of the HVAC system in response to changes in real-time data by integrating disturbance predictions from the time-series underlying model with real-time feedback (collected real-time data).

[0020] Some embodiments are based on the further recognition that real-time data can be used to adapt the probabilistic feedback controller to real-world conditions by aligning predictions with actual results. In particular, the feedback mechanism is configured to update the probabilistic feedback controller based on the difference between the predicted data and the real-time data.

[0021] Accordingly, one embodiment discloses a feedback control system for optimizing the performance of a grid-interactive building (GIB) system configured to regulate the indoor environment of a building. The feedback control system comprises a time-series underlying model configured to predict disturbances affecting the building's energy consumption and a probabilistic feedback controller, the probabilistic feedback controller configured to determine a control input to the GIB system by evaluating a number of control actions to the GIB system based on the predicted disturbances affecting the building's energy consumption, the control input maximizing the likelihood of achieving desired indoor environmental conditions while minimizing energy consumption. The feedback control system further comprises a feedback mechanism configured to collect real-time data including indoor environmental conditions, actual energy consumption, and external factors, to fine-tune the time-series underlying model using the collected real-time data, and to update the probabilistic feedback controller based on the difference between the predicted data and the collected real-time data.

[0022] Accordingly, another embodiment discloses a method for optimizing the performance of a grid-interactive building (GIB) system configured to regulate the indoor environment of a building. The method includes a time-series-based model predicting disturbances affecting the building's energy consumption, and a probabilistic feedback controller determining control inputs to the GIB system by evaluating a number of control actions for the GIB system based on the predicted disturbances affecting the building's energy consumption, the control inputs maximizing the likelihood of achieving desired indoor environmental conditions while minimizing energy consumption. The method further includes a feedback mechanism collecting real-time data including indoor environmental conditions, actual energy consumption, and external factors, the feedback mechanism fine-tuning the time-series-based model using the collected real-time data, and the feedback mechanism updating the probabilistic feedback controller based on the difference between the predicted data and the collected real-time data.

[0023] Thus, yet another embodiment discloses a non - transient computer - readable storage medium having a program executable by a processor to perform a method for optimizing the performance of a grid - interactive building (GIB) system configured to condition an indoor environment of a building. The method includes a time - series - based model predicting disturbances that affect the energy consumption of the building, and a probabilistic feedback controller determining control inputs for the GIB system by evaluating a plurality of control actions for the GIB system based on the predicted disturbances that affect the energy consumption of the building, the control inputs maximizing the likelihood of achieving desired indoor environmental conditions while minimizing energy consumption. The method further includes a feedback mechanism collecting real - time data including indoor environmental conditions, actual energy consumption, and external factors, the feedback mechanism fine - tuning the time - series - based model using the collected real - time data, and the feedback mechanism updating the probabilistic feedback controller based on the difference between the predicted data and the collected real - time data.

[0024] The embodiments disclosed herein are further described with reference to the accompanying drawings. The drawings shown are not necessarily to scale; rather, emphasis is generally placed upon illustrating the principles of the embodiments of the present disclosure.

Brief Description of the Drawings

[0025] [Figure 1] A schematic diagram of an overview of disturbance - aware control employed by some embodiments is shown. [Figure 2] A schematic diagram of training and estimation stages employing a generative artificial intelligence (AI) for disturbance signal generation according to some embodiments is shown. [Figure 3] A schematic diagram of a conditional variational auto - encoder (CVAE) employed by some embodiments to provide a mapping between a latent space and an original space is shown. [Figure 4]Schematic diagrams of conditional probability distributions as specific examples of latent representations according to some embodiments are shown. [Figure 5] It is a diagram showing a flowchart of a method for obtaining a conditional probability distribution of a latent representation of noise according to some embodiments. [Figure 6] Schematic diagrams of embodiments using sigma points derived from an estimation score to generate conditional probabilities are shown. [Figure 7] It is a diagram showing a flowchart of a method adopting the principle described with respect to FIG. 6. [Figure 8] It is a diagram showing an example of a system with uncertainty connected to a stochastic model predictive controller (SMPC) via a noise estimator according to some embodiments. [Figure 9] Diagrams showing methods implemented by the SMPC of FIG. 8 according to some embodiments are shown. [Figure 10A] It is a diagram showing a flowchart of an example of a process according to some embodiments. [Figure 10B] It is a diagram showing a flowchart of an example of a process according to some embodiments. [Figure 11] It is a diagram showing pseudocode for executing SMPC using a scenario tree according to some embodiments. [Figure 12] It is a diagram showing pseudocode for an implementation scenario tree as a specific example of the SMPC of FIG. 11 for building energy control according to some embodiments. [Figure 13A] It is a diagram showing a feedback control system for optimizing the performance of a grid interactive building (GIB) system according to some embodiments of the present disclosure. [Figure 13B] It is a diagram showing fine-tuning of a time-series based model according to some embodiments of the present disclosure. [Figure 14] It is a diagram showing training of a time-series based model according to an embodiment of the present disclosure. [Figure 15] It is a diagram showing an overview of reinforcement learning according to an embodiment of the present disclosure. [Figure 16] This figure shows the joint optimization of a time-series-based model and a probabilistic feedback controller according to some embodiments of the present disclosure. [Figure 17] A block diagram is shown for fine-tuning a time-series underlying model using low-rank approximations, according to some embodiments of this disclosure. [Figure 18] This figure shows the application of a low-rank approximation to a specific layer of a time-series-based model according to some embodiments of this disclosure. [Figure 19] This figure shows a hierarchical training process according to some embodiments of the present disclosure. [Figure 20] The following are block diagrams for predicting disturbances in unmonitored zones of a building, according to some embodiments of this disclosure. [Figure 21] This figure shows a feedback control system including a multivariate quantile predictor (MQF2) according to some embodiments of the present disclosure. [Figure 22A] A block diagram of a time-fusion transformer (TFT) model forming an MQF2, according to some embodiments of this disclosure, is shown. [Figure 22B] This figure shows a modified input to a time-series-based model according to some embodiments of this disclosure. [Figure 23] This is a schematic diagram illustrating, as a non-limiting example, a computing device for realizing the method and system disclosed herein. [Modes for carrying out the invention]

[0026] The following description includes numerous specific details for the purpose of explaining the disclosure in order to fully understand it. However, it will be apparent to those skilled in the art that the disclosure can be implemented without these specific details. In other examples, the apparatus and methods are shown in block diagram form solely for the purpose of avoiding obscurity of the disclosure.

[0027] The terms “for example,” “for instance,” and “such as,” as used herein and in the claims, as well as the verbs “comprising,” “having,” and “including,” and each of these verbs in other forms, should be interpreted as open-ended, meaning that when used with an enumeration of one or more components or other items, the enumeration should not be considered to exclude any further components or items. The term “based on” means based at least partially. Furthermore, it should be understood that the style and terminology used herein are for illustrative purposes only and should not be considered restrictive. Any headings used herein are for convenience only and have no legal or restrictive effect.

[0028]

number

[0029] For example, the mechanical system may be a thermodynamic system that regulates energy within a building, and model f can describe the thermodynamics of the air-conditioned zone within the building, with the system state x including indoor air temperature, interior wall surface temperature, and exterior wall core temperature. The control input u could be the net heating and cooling capacity of the heat pump, and the disturbance input vector w could be one or a combination of the following: outside air temperature, solar radiation, airflow, and internal heat load due to the presence of occupants and heat-generating equipment.

[0030] The objective of the disturbance-aware control algorithm 100 is to achieve desired building operating conditions 109 based on state estimates, sensor measurements, and disturbance input predictions. These predictions are generated using a generative AI 110 trained on data including disturbance inputs 111 measured in real time from the mechanical system 101, or disturbance inputs previously stored in a database 115 collected offline from either the historical building energy system 101 or an alternative data source.

[0031] Figure 2 shows schematic diagrams of the training and estimation stages employing generative AI for disturbance signal generation according to several embodiments. Some embodiments are based on the recognition that measurements of time-series data indicating disturbances have at least some unknown relationships in the time domain. An example of such relationships can be observed in sensors measuring the operation of a power plant where future load fluctuations depend on the current load value. Some embodiments are based on the recognition that it is difficult to determine unknown relationships because the measurements in the sensor's original data space are noisy and the unknown relationships involve complex nonlinear transformations. For example, in the case of a power plant, the thermodynamic relationships in the power plant are complex and require extensive domain knowledge to elucidate. Such complex interdependencies make the recovery of relationships in the original data space unreliable. Therefore, the assumption of a parameterized structure of the distribution that captures this relationship is also unreliable.

[0032] Some embodiments are based on the understanding that relationships between sensor measurements can be found through efficient encoding of the measurements, because the encoding method is used to find a dimensionality-reduced embedding of the data that summarizes those relationships in the original data space. In addition, some embodiments are based on the understanding that if the dimensionality-reduced embedding of the measurements can better represent the relationships between the measurements, then inaccurate assumptions about the structure of the probabilistic model of the disturbance in the measurement domain, for example, the original domain, can become accurate in the domain of the dimensionality-reduced embedding of the measurements.

[0033] Therefore, during the training phase of a control process performed offline, some embodiments learn a latent space of a dimensionality-reduced embedding of the original measurements of the disturbance (210), and the latent representation of the disturbance can be modeled on a parameterized distribution with sufficient precision for control applications. As described below, embodiments learn a deep generative decoder model that defines a mapping from the latent space of the latent representation of the time-series values ​​of the disturbance affecting the machine system over a time horizon to, for example, the measured space of the original disturbance.

[0034] In particular, the latent space encodes the time series values ​​of disturbances that affect the machine system over the time horizon. For example, during training, the time horizon may be 24 hours, and the time series values ​​of disturbances could be measured disturbances that affect the system over a 24-hour period. Therefore, in this example, each sample in the latent space encodes a disturbance trajectory with a length of 24 hours.

[0035] During online control, partial observations of disturbances are collected (220) and measured, for example, the disturbance is measured in the last hour. However, the prediction horizon of the SMPC may be longer than one hour. For example, in this example, it could be several hours or up to 24 hours. Therefore, it is necessary to predict the remaining unknown disturbances conditional on the partial observations of the disturbances. However, considering the latent space, instead of attempting to find the unstructured distribution of the disturbances conditional on the partial observations in the original measurement space, the embodiment finds the parameterized distribution of the latent representation of the disturbances in the latent space conditional on the partial observations in the original measurement domain (230). As mentioned above, due to the nature of the latent space, such estimations are more accurate. The conditional distribution allows sampling of the disturbance signal in the latent space (240) and allows using the decoded latent samples for the disturbance-aware probabilistic control (250).

[0036] To perform such estimations, some embodiments use a deep generative decoder model that uses a mapping between the latent space and the original space determined offline. Some embodiments are based on the understanding that an autoencoder can find such an efficient mapping unsupervised. An autoencoder is a type of artificial neural network used to learn efficient data coding unsupervised. An autoencoder includes an encoder and a decoder. The encoder encodes input data from the original data space into a latent space represented by a vector of numbers "h". The decoder decodes the encoding from the latent space into an estimate of the input data, i.e., reconstructs the input data. In other words, the encoder and / or decoder provide a mapping between the data in the original data space and its latent space representation. To this end, the autoencoder determines an efficient latent space suitable for capturing the relationships between different instances of the input data.

[0037] By extending the principles of autoencoders to deep generative models such as variational autoencoders (VAEs) or conditional variational autoencoders (CVAEs), it is possible to provide a representational automation method for learning the distribution in the latent space from data measured in the original space.

[0038]

number

[0039]

number

[0040]

number

[0041]

number

[0042]

number

[0043] The variational posterior can be considered an encoder that induces a probabilistic map from W to latent representation z, conditioned on c. The generative model can be considered a decoder that recovers the likelihood for W, conditioned on c, from the sampled latent representation z. This decoder can also be parameterized as a conditional Gaussian, where the mean vector is a parametric function of (z,c) and the covariance is the identity matrix. This simplifies the first term of the ELBO to a substantially negative reconstruction loss, i.e., a shift scale of the mean squared error (MSE).

[0044]

number

[0045] However, the deep generative decoder model 305 is trained offline from measured disturbance training data without considering the currently partially observed disturbances acting on the machine system. To address partial observations, the embodiment uses the deep generative decoder model to find a conditional probability distribution of the latent representation of the disturbance conditioned on partial observations of the disturbance.

[0046] Figure 4 shows a schematic diagram of a conditional probability distribution as a concrete example of a latent representation according to several embodiments. The conditional probability distribution 410 of the latent representation of the disturbance is determined based on a comparison between the corresponding portion of the set of latent representations decoded by the deep generative decoder model and the partial observation of the disturbance. For example, in the example in Figure 4, decoding of the latent sample from area 420 is more likely to fit the partial observation than decoding of the latent sample from area 430.

[0047] In some embodiments, a conditional distribution 410 is sampled (460) to generate a latent sample 450 and its probability 440, representing a partial observation. The decoding of the latent sample 450 and its probability 440 is used by the SMPC for probabilistic control.

[0048] Different embodiments use different techniques to determine the conditional distribution 410. For example, some embodiments determine the conditional probability distribution of the latent representation of the disturbance based on a comparison of a corresponding portion of the set of latent representations decoded by a deep generative decoder model with a partial observation of the disturbance.

[0049] Figure 5 shows a flowchart of a method for determining the conditional probability distribution of a latent representation of a disturbance according to several embodiments. This method includes generating a set of latent samples by sampling the probability distribution of the latent representation 510, and obtaining a set of time-series values ​​of the disturbance over the time horizon by decoding each of the latent samples using a deep generative decoder model 520, where each of the time-series values ​​of the disturbance includes values ​​over the observed portion of the time horizon, and the method further includes generating a set of scores 530 by comparing the values ​​over the observed portion of the time horizon in the obtained set of time-series values ​​of the disturbance with partial observations of the disturbance.

[0050] These scores are used to construct a conditional distribution 410. For example, in some embodiments, sampling 510, decoding 520, and comparison 530 are repeated until a termination condition is met in order to reduce the error between the values ​​across the observed portion of the time horizon in the obtained set of time-series values ​​of the disturbance and the partial observation of the disturbance. In this way, an area of ​​the latent space with a high probability of fitting the partial observation of the disturbance can be identified. The observed error is used to estimate the conditional probability for different samples.

[0051]

number

[0052]

number

[0053]

number

[0054]

number

[0055]

number

[0056] Next, these latent samples, along with their probabilities, are sent to the decoder to generate conditionally sampled predictions of the still-unknown portion of the disturbance sequence, and subsequently weights 411 sent to the SMPC via the predicted dynamics.

[0057] Figure 7 shows a flowchart of a method employing the principle described in relation to Figure 6. This method involves approximating a conditional probability distribution of the latent representation of the disturbance, conditioned on partial observation of the disturbance, using a kernel density estimate (KDE) of a set of scores.710 For example, the conditional probability distribution is approximated as a set of sigma point samples on the KDE of the set of scores. The method then uses the set of sigma point samples as a set of latent samples of time-series values ​​of disturbances affecting the machine system over a time horizon to generate a set of disturbance scenarios affecting the machine system over a time period (720), sends the set of disturbance scenarios with corresponding probabilities of the set of sigma point samples to a predictive controller (730), and generates control commands by optimizing the cost function of the set of scenarios weighted by the corresponding probabilities.

[0058] Figure 8 shows a system 820 as an example with uncertainty 825, connected to a stochastic model predictive controller (SMPC) 810 via a disturbance estimator 831, according to one embodiment. The SMPC is programmed according to a dynamic model 840, i.e., a control model of the system. The model may be a set of equations that describe the state of the system 820 and the changes in its output 803 over a period of time, as a function of current and previous inputs 811 and previous outputs 803. The model may include constraints 842 that represent the physical and operational limitations of the system. During operation, the controller receives a command 801 that indicates the desired behavior of the system. The command may be, for example, a motion command. In response to receiving a command 801, the controller generates a control signal 811 that serves as an input to the mechanical system 820 affected by the disturbance 825. In response to this input, the system updates its output 803. Based on the measured output 803 of the system and an AI deep generative decoder model 850, the estimator 831 predicts the disturbance 825 and its uncertainty (821). These estimated values ​​821 are sent to the controller 810.

[0059] The mechanical system 820 referred to herein may be any machine or device that is controlled by specific operational input signals 811 (inputs), which may be associated with physical quantities such as voltage, pressure, force, and torque, and returns several controlled output signals 803 (outputs), which may be associated with physical quantities such as current, flow rate, velocity, and position, indicating a transition of the system's state from a previous state to a current state. The output values ​​are partly related to the system's previous output values ​​and partly related to previous and current input values. Dependencies on previous inputs and previous outputs are encoded by the system's state. The operation of the system, for example, the motion of the system's components, may include a sequence of output values ​​generated by the system after the application of some input value.

[0060] Uncertain disturbances 825 can be any time-varying signals, any unmodeled dynamics, or any uncertainties in physical quantities, including any external disturbances, forces, or torques acting on system 820, or uncertain coefficients and parameters in control model equations describing the physical behavior of the real system 820, such as uncertain coefficients of friction, friction functions, mass of an object, center of gravity of the system, or uncertain coefficients and parameters in control model equations describing the physical behavior of the real system 820. For example, in some implementations, the SMPC 810 uses a simplified control model 840 to reduce the complexity of controller calculations, or because some of the physical behavior is too complex to be modeled by the first principle, resulting in a large amount of physical behavior in the mechanical system remaining unmodeled. Such simplified modeling can cause or contribute to uncertainty 825. Time-independent uncertainty can be estimated or learned online or offline as part of the state and parameter estimator 831.

[0061] In various embodiments, the estimator 831 is an online estimator that determines the confidence level of uncertain disturbances 825 and / or estimated uncertainties in real time, i.e., during control of the system 820. Since uncertainties 825 may change over time and depend on the control inputs and the system's response to such control inputs, some embodiments thus improve the accuracy of the estimation of uncertainties 825 with respect to the accuracy of the offline estimation of uncertainties.

[0062] The control model 840 may include a dynamic model that defines the dynamics of the system 820. The control model 840 of the mechanical system 820 may include a set of mathematical equations that describe how the system output changes over time as a function of the current input and previous inputs and outputs. The system state is, along with the system model and future inputs, a set of information that generally changes over time, such as the current inputs and outputs and a suitable subset of previous inputs and outputs, which can uniquely define the future motion of the system. The mechanical system 820 may be subject to physical and specification constraints 842 that limit the range in which the outputs, inputs, and possibly the system state can operate. In various embodiments, the control model of the system includes a function of the system dynamics with parameters that have uncertainty 825. In this way, uncertainty acting on the system 820 can be captured by the model 840.

[0063] The controller 810 receives the estimated state of the system 821 and a desired motion command 801 at a fixed or variable control period sampling interval, and uses this information to determine an input for operating the system, such as a control signal 811. This can be implemented in hardware or as a software program executed on a processor, such as a microprocessor.

[0064] The estimator 831 receives the system's output 803 at a fixed or variable controlled sampling interval and uses new and previous output measurements to determine the estimated disturbance of the system 820 and its uncertainty 821. This can be implemented in hardware or as a software program running on a processor that is either the same or a different processor as the controller 810.

[0065] Figure 9 shows a diagram of the method implemented by the SMPC of Figure 8 in several embodiments. Considering a deep generative decoder model that defines a mapping from the latent space of the latent representation of the time-series values ​​of disturbances affecting the machine system over the time horizon to the measured space of partial observations of the disturbance, the method identifies a subspace of latent variables that is most likely to have generated the disturbance signal measured up to the present time using a conditional input (910). This identification 910 can be implemented as a conditional probability distribution of the latent representation of the disturbance conditioned on partial observations of the disturbance.

[0066] This method uses an identified subspace of latent variables and a deep generative decoder model to predict future disturbance inputs based on the identified latent subspace (920), and uses this prediction to calculate the cost function and statistics of stochastic constraint violations across different realizations of the disturbance, as described with respect to Figure 8, for example (930). For example, the SMPC determines control commands by optimizing the cost function across a predicted horizon that includes the observed and unobserved portions of the time horizon, and the time series values ​​of disturbances affecting the machine system across the predicted horizon include partial observations of the disturbance, complemented by some of the predicted values ​​of the disturbance in the unobserved portion of the time horizon. In particular, in different implementations, the predicted horizon is shorter than or equal to the time horizon of the decoded disturbance.

[0067] Next, the SMPC performs an iterative optimization procedure to select the best sequence of input values ​​that minimizes the cost of being stochastically constrained across the prediction horizon (940), and sends a portion of the optimized control inputs in the sequence to the actuators of the mechanical system (950).

[0068] Figures 10A and 10B show a flowchart of process 1000 as an example. In some implementations, one or more process blocks in Figures 10A and 10B may be executed by a processor. As shown in Figures 10A and 10B, process 1000 may include collecting partial observations of disturbances affecting the operation of a machine system across the observed portion of the time horizon (block 1002). For example, as described above, the processor may collect partial observations of disturbances affecting the operation of a machine system across the observed portion of the time horizon. Also, as shown in Figures 10A and 10B, process 1000 may include collecting a deep generative decoder model that defines a mapping from the latent space of latent representations of time-series values ​​of disturbances affecting a machine system across the time horizon to the measured space of partial observations of disturbances (block 1004). For example, as described above, the processor may collect a deep generative decoder model that defines a mapping from the latent space of latent representations of time-series values ​​of disturbances affecting the machine system over the time horizon to the measurement space of partial observations of the disturbances.

[0069] As further shown in Figures 10A and 10B, process 1000 may include using a deep generative decoder model to find a conditional probability distribution of the latent representation of the disturbance conditioned on partial observation of the disturbance (block 1006). For example, as described above, the processor may use a deep generative decoder model to find a conditional probability distribution of the latent representation of the disturbance conditioned on partial observation of the disturbance.

[0070] Furthermore, as shown in Figures 10A to 10B, process 1000 may include sampling the conditional probability distribution of the latent representation to generate a latent sample of time-series values ​​of disturbances affecting the machine system over the time horizon (block 1008). For example, as described above, the processor may sample the conditional probability distribution of the latent representation to generate a latent sample of time-series values ​​of disturbances affecting the machine system over the time horizon.

[0071] As further shown in Figures 10A and 10B, process 1000 may include decoding latent samples using a deep generative decoder model to generate predicted values ​​of disturbances acting on the system within the time horizon at the probability of the latent samples relative to the conditional probability distribution of the latent representation (block 1010). For example, as described above, the processor may decode latent samples using a deep generative decoder model to generate predicted values ​​of disturbances acting on the system within the time horizon at the probability of the latent samples relative to the conditional probability distribution of the latent representation.

[0072] Furthermore, as shown in Figures 10A to 10B, process 1000 may include controlling the machine system using a predictive controller that determines control commands to change the operating state of the machine system using at least some probabilities of predicted disturbances (block 1012). For example, as described above, the processor may control the machine system using a predictive controller that determines control commands to change the operating state of the machine system using at least some probabilities of predicted disturbances.

[0073] Figures 10A-10B show examples of blocks in process 1000, but in some implementations, process 1000 may include additional blocks, fewer blocks, different blocks, or blocks arranged differently from those shown in Figure 10. In addition, or instead, two or more blocks of process 1000 may run in parallel.

[0074] Figure 11 shows pseudocode for performing an SMPC using a scenario tree according to several embodiments. An SMPC is an extension of an MPC that describes uncertainty in a system by modeling it as a probabilistic process, while a scenario tree is a tool used to represent and handle these uncertainties in a probabilistic MPC. Several embodiments are based on the recognition that scenario trees can be conveniently used to deal with different predictions of disturbances sampled with different probabilities about a conditional distribution.

[0075] As will be readily apparent to those skilled in the art, the scenario tree is constructed to represent different possible realizations of an uncertain variable across the prediction horizon. Each branch of the tree corresponds to a specific scenario or realization of the uncertainty. Nodes within the scenario tree represent decision points, such as control inputs or sampling points for prediction steps. At each node, the system faces a decision, and the tree branches based on different possible outcomes. Each branch of the scenario tree is associated with probability weights that reflect the likelihood of that particular scenario occurring. As described above, these probabilities are estimated based on a conditional distribution.

[0076] In probabilistic MPC, the objective function is defined as the expected cost across all possible scenarios, taking into account the probability of each scenario. This involves weighting the cost associated with each scenario by its probability of occurrence. The optimization problem in probabilistic MPC involves finding the control input that minimizes the expected cost across the entire scenario tree. This leads to a more robust controller that performs well on average across different possible outcomes.

[0077] Probabilistic MPCs typically employ a backward-horizon control strategy. At each time step, the controller solves an optimization problem across the current scenario tree, implements a first set of control inputs, and then updates the scenario tree based on new measurements. Over time, actual system behavior is observed, and the scenario tree may be updated to incorporate the new information. This adaptive approach helps the controller become more accurate over time.

[0078] Scenario trees provide a structured method for making decisions while dealing with uncertainty in a probabilistic environment. They allow MPC controllers to explicitly consider multiple possible disturbance scenarios, making control strategies more robust and enabling them to handle real-world uncertainty.

[0079]

number

[0080]

number

[0081] Assuming that the learned distribution can generate all possible scenarios, it will also include all scenarios in a tree with an arbitrarily long, robust horizon, even without explicitly defining branching and transition probabilities. Thus, taking a subset of predictions can be likened to a pruned scenario tree.

[0082]

number

[0083]

number

[0084]

number

[0085] As a result, SMPC can solve this scenario tree optimal control problem using various iterative optimization methods and transmit a portion of the optimal control solution to the building energy system. [Optimizing the performance of grid-interactive building systems under the influence of disturbances using time-series foundational models]

[0086] Figure 13A shows a feedback control system 1301 for optimizing the performance of a grid-interactive building (GIB) system 1303 according to some embodiments of the present disclosure. The GIB system 1303 includes one or more of the following: a heating, ventilation, and air conditioning (HVAC) system, a renewable energy source, and a power grid. The HVAC system, the renewable energy source, and the power grid are electrically interconnected.

[0087] HVAC systems are installed within a building and configured to regulate the building's indoor environment. For example, an HVAC system is configured to maintain the indoor temperature and / or humidity within a desired range or at a desired value. The indoor environment of a building refers to the space of the building's rooms or floors. In one embodiment, the HVAC system is powered by electrical energy from the power grid. In addition, the building is supplied with electrical energy from the power grid. The power grid is a network of transmission and distribution systems that send electricity from power plants to homes, businesses, and other users. It is the infrastructure that enables the generation, transmission, and distribution of electrical energy over long distances.

[0088] The objective of some embodiments is to optimize the performance of the GIB system 1303. For example, the objective of some embodiments is to minimize the electrical energy supplied from the power grid by utilizing renewable energy sources. Renewable energy sources include on-site renewable energy sources and on-site energy storage systems, such as photovoltaic systems and batteries associated with the building. Also, the objective of some embodiments is to minimize the energy consumption of the HVAC system and the building. In addition, the objective of some embodiments is to minimize the energy consumption of the HVAC system while maintaining the indoor environment temperature and humidity at desired values.

[0089] To achieve such objectives, embodiments of the present disclosure provide a feedback control system 1301. The feedback control system 1301 is communicatively coupled to a GIB system 1303. In some other embodiments, the feedback control system 1301 is integrated into the GIB system 1303. The feedback control system 1301 includes a processor 1305, a memory 1307, a time-series underlying model 1309, a probabilistic feedback controller 1311, and a feedback mechanism 1313. The processor 1305 may be a single-core processor, a multi-core processor, a computing cluster, or any number of other configurations. The memory 1307 may include random-access memory (RAM), read-only memory (ROM), flash memory, or any other suitable memory system. In addition, in some embodiments, the memory 1307 may be implemented using a hard drive, an optical drive, a thumb drive, an array of drives, or any combination thereof. In some embodiments, the time-series-based model 1309, the probabilistic feedback controller 1311, and the feedback mechanism 1313 are modules of the feedback control system 1301, which is executed by the processor 1305.

[0090] The time-series underlying model 1309 is a pre-trained model configured to predict disturbances affecting the building's energy consumption. These disturbances include external factors describing weather conditions, occupancy patterns in the indoor environment, and external temperature fluctuations. The predicted disturbances are input to the probabilistic feedback controller 1311.

[0091] The probabilistic feedback controller 1311 is configured to determine control inputs to the GIB system 1303 by evaluating multiple control actions for the GIB system based on disturbances affecting the predicted energy consumption of the building. The control inputs are determined to maximize the likelihood of achieving desired indoor environmental conditions while minimizing energy consumption. Desired indoor environmental conditions include desired temperature and humidity of the indoor environment. The control inputs include, for example, the compressor speed of the HVAC system, the position of the expansion valve, and the speeds of the indoor and outdoor fans of the HVAC system. The probabilistic feedback controller 1311 is further configured to generate control commands for the actuators of the HVAC system based on the control inputs and to control the actuators of the HVAC system according to the control commands to operate the HVAC system according to the control inputs.

[0092] The probabilistic feedback controller 1311 controls the HVAC system according to control inputs determined by considering predicted uncertainties, thereby accurately achieving the desired temperature and humidity of the indoor environment while minimizing energy consumption of the HVAC system and the building. This optimizes the performance of the GIB system 1303.

[0093] Some embodiments are based on the understanding that the predictive capability of the time series-based model 1309 can be improved by fine-tuning it based on real-time data. To this end, some embodiments provide a feedback mechanism 1313 configured to fine-tune the time series-based model 1309 based on real-time data.

[0094] Figure 13B shows a fine-tuning of the time-series-based model 1309 according to some embodiments of the present disclosure. A probabilistic feedback mechanism 1313 is configured to collect real-time data 1315. The real-time data 1315 includes indoor environmental conditions, actual energy consumption by the HVAC system, and external factors. Indoor environmental conditions include temperature, humidity, and carbon dioxide (CO2) levels in the indoor environment. External factors include ambient temperature, ambient humidity, and solar radiation. In addition, in some embodiments, the real-time data 1315 includes inputs targeting occupants, such as equipment energy usage.

[0095] Furthermore, the feedback mechanism 1313 is configured to fine-tune the time-series underlying model 1309 based on real-time data 1315. By fine-tuning the time-series underlying model 1309 based on real-time data 1315, the accuracy of disturbance prediction is improved. In some embodiments, the feedback mechanism 1313 periodically collects real-time data 1315 and periodically updates the time-series underlying model 1309. In this way, the feedback control system 1301 dynamically optimizes the performance of the HVAC system in response to changes in real-time data by integrating disturbance predictions from the time-series underlying model 1309 with real-time feedback (collected real-time data).

[0096] Some embodiments are based on the further recognition that real-time data 1315 can be used to adapt the probabilistic feedback controller 1311 to real-world conditions by aligning predictions with actual results. In particular, the feedback mechanism 1313 is configured to update the probabilistic feedback controller 1313 based on the difference between the predicted data and the real-time data 1315. The predicted data includes the energy consumption of the HVAC system and / or building predicted by the probabilistic feedback controller 1313.

[0097] In one embodiment, a probabilistic feedback controller 1311 uses Bayesian estimation to evaluate multiple control actions for a GIB system 1303. Bayesian estimation is a statistical method that allows updating the probability of a hypothesis or event based on new data. Bayesian estimation is based on Bayes' theorem, which describes the probability of a hypothesis considering some observed data. Bayesian estimation involves starting with an prior belief about the hypothesis (called the prior probability) and then updating the prior belief based on new data (called the likelihood) to obtain a more refined belief (the posterior probability).

[0098] In another embodiment, the probabilistic feedback controller 1311 is a probabilistic optimization technique for evaluating multiple control actions on the GIB system 1303. A probabilistic optimization technique is a type of optimization technique that involves randomness or probabilistic elements in its process. These techniques are used to solve optimization problems in which the objective function or multiple constraints are affected by uncertainty, noise, or randomness, or when a deterministic evaluation of the objective function is not feasible due to factors such as complex computation, high computational cost, or incomplete data. Examples of probabilistic optimization techniques include stochastic gradient descent (SGD), simulated annealing, and stochastic programming.

[0099] Figure 14 shows the training of a time-series foundational model 1309 according to one embodiment of the present disclosure. A time-series foundational model 1309 is a type of machine learning model specifically designed to handle, understand, and make predictions from time-series data. Time-series data typically consists of a sequence of data points ordered relative to time and is used to model phenomena that evolve over time. In this context, a foundational model is a large, pre-trained, general-purpose model designed to capture complex patterns in time-series data. A foundational model can be fine-tuned or adapted to a specific task using time-series data as input.

[0100] In this disclosure, a time series-based model 1403 is pre-trained on a general-purpose dataset 1401 to obtain a pre-trained time series-based model 1405. The general-purpose dataset 1401 contains a large and diverse set of time series data used to help the time series-based model learn common temporal patterns, structures, and relationships across different domains. The general-purpose dataset 1401 is large enough to capture a wide variety of temporal patterns and behaviors, allowing the time series-based model 1403 to be generalized across multiple domains and tasks. Furthermore, the diverse set of time series data includes time series data from different sources and industries (e.g., finance, healthcare, weather, retail, energy) to ensure that the time series-based model 1403 can handle different types of time series phenomena such as seasonality, trends, and noise.

[0101] A time-series base model 1309 is obtained by fine-tuning a pre-trained time-series base model 1405 to fit the building's energy consumption patterns using building-specific historical data 1407. In particular, the fine-tuning updates the pre-trained time-series base model 1405 to capture specific features of the building's energy consumption, such as daily cycles, seasonal trends, occupancy patterns, and responses to environmental factors (e.g., weather, time of day). The building-specific historical data 1407 includes data collected directly from energy meters or sensors within the building, showing total energy use over time (typically hourly or minutely data). The collected data includes historical data on electricity, heating, cooling, or even gas consumption. In addition, in some embodiments, the building-specific historical data 1407 includes occupancy patterns that can help the time-series base model 1309 understand the correlation between the energy consumption of the HVAC system and the number of people in the building.

[0102] In some embodiments, the feedback mechanism 1313 is configured to update a time-series-based model 1309 based on real-time data 1315 using a reinforcement learning framework. Figure 15 is a diagram illustrating an overview of reinforcement learning according to an embodiment of the present disclosure. Reinforcement learning is a learning framework that deals with sequential decision problems in which an “agent” 1530 or decision-maker learns policies and optimizes long-term rewards by interacting with an (unknown) environment 1510. At each time step, the reinforcement learning agent receives evaluation feedback (called reward or cost) 1550 about the performance of its action 1540, along with observations of the environment, which allows it to improve (maximize or minimize) the performance of subsequent actions.

[0103] In some other embodiments, the probabilistic feedback controller 1311 and the time-series-based model 1309 are optimized jointly. Figure 16 shows the joint optimization of the time-series-based model 1309 and the probabilistic feedback controller 1311 in some embodiments of the present disclosure. The processor 1305 is configured to jointly optimize the time-series-based model 1309 and the probabilistic feedback controller 1311 (1601). Such joint optimization 1601 minimizes the energy consumption of a building while maintaining indoor environmental conditions within a predetermined range. For example, indoor environmental conditions include indoor temperature and humidity. The indoor temperature is maintained within a predetermined temperature range. Similarly, the indoor humidity is maintained within a predetermined humidity range. The predetermined range is defined by the user.

[0104] In some embodiments, the time series-based model 1309 is fine-tuned using a dimensionality reduction technique based on low-rank approximation. Low-rank approximation is a technique used in numerical and data science applications to approximate a matrix or a high-dimensional object with a lower-rank matrix. This is typically done to reduce the complexity of the data representation while preserving essential information, making computations more efficient, or reducing memory requirements. In this disclosure, the time series-based model 1309 is fine-tuned using low-rank approximation to reduce the complexity of the time series-based model 1309 while preserving essential predictive features.

[0105] Figure 17 shows a block diagram for fine-tuning the time-series underlying model 1309 using low-rank approximation, according to some embodiments of the present disclosure. Processor 1305 is configured to fine-tune the time-series underlying model 1309 using low-rank approximation. In block 1701, low-rank approximation includes identifying low-rank subspaces within the feature representation of the time-series underlying model 1309. In the context of the time-series underlying model 1309, a low-rank subspace is an underlying dimensionality reduction structure that captures the most important patterns or trends in the time-series data while ignoring less important noise-like components. Low-rank subspaces represent a method of compressing or approximating time-series data, making its processing more efficient, particularly when dealing with large datasets or high-dimensional problems.

[0106] In block 1703, the low-rank approximation includes updating the low-rank subspace during fine-tuning. In block 1705, the low-rank approximation includes maintaining fixed, pre-trained high-rank parameters to preserve the general knowledge of the time-series underlying model 1309. The general knowledge of the time-series underlying model 1309 corresponds to the predictive power that the time-series underlying model 1309 acquires after being trained on the general-purpose dataset 1401. Since only the low-rank subspace is updated during fine-tuning, the low-rank approximation accelerates the fine-tuning of the time-series underlying model 1309 by enabling optimization across a reduced parameter space, thereby facilitating real-time or near-real-time model adaptation.

[0107] In some embodiments, the low-rank approximation is selectively applied to specific layers or components of the time-series-based model 1309.

[0108] Figure 18 illustrates the application of a low-rank approximation to a specific layer of a time-series-based model 1309 in some embodiments of the present disclosure. The time-series-based model 1309 includes various layers such as layer 1 1801, layer 2 1803, layer 3 1805, ..., layer N 1807. In one embodiment, the processor 1305 is configured to apply the low-rank approximation to a specific layer, for example, layer 2 1803 and layer 3 1805. By applying the low-rank approximation to a specific layer of the time-series-based model 1309, the time-series-based model 1309 is fitted to the building's energy consumption patterns while preserving the pre-trained representation. Furthermore, the low-rank approximation reduces the number of trainable parameters in the time-series-based model 1309, enabling efficient fine-tuning to building-specific historical data 1407 with limited computational resources.

[0109] In some embodiments, the low-rank approximation is dynamically adjusted during fine-tuning based on the available computational resources of the feedback control system 1301 or the size of the building-specific historical data 1407. Such dynamic adjustment ensures that the low-rank approximation is applied to the fine-tuning of the time-series underlying model 1309.

[0110] Several methods are used to train the time-series-based model 1309. For example, in one embodiment, the time-series-based model 1309 is trained hierarchically by using a hierarchical training process. In another embodiment, the time-series-based model 1309 is trained using a federative learning method.

[0111] Figure 19 shows a hierarchical training process according to some embodiments of the present disclosure. First, data is aggregated from multiple buildings, such as Building 1 1901, Building 2 1903, ..., Building N 1905, to form a comprehensive building dataset 1907. The data from each building includes data collected directly from energy meters or sensors within each building, showing total energy consumption over time (typically hourly or minutely). In addition, in some embodiments, the data from each building includes occupancy patterns for each building, which can help the model understand the correlation between the energy consumption of the HVAC system and the number of people in the building.

[0112] Furthermore, the time-series underlying model 1309 is trained using the comprehensive building dataset 1907 to learn global patterns of building energy consumption. The time-series underlying model 1309, trained on the comprehensive building dataset 1907, is then fine-tuned using building-specific historical data 1407 to fit the building energy consumption patterns.

[0113] In some cases, the building-specific historical data 1407 for a building may be sparse or incomplete. Fine-tuning the time-series-based model 1309 using such sparse or incomplete data may result in a partial or inadequate fine-tuning of the time-series-based model 1309. To mitigate this issue, the time-series-based model 1309 is fine-tuned using data from other buildings with similar structural or operational characteristics to predict disturbances.

[0114] In some embodiments, the time-series underlying model 1309 predicts disturbances in unmonitored zones of a building.

[0115] Figure 20 shows a block diagram for predicting disturbances in unmonitored zones of a building according to some embodiments of the present disclosure. In block 2001, the time-series underlying model 1309 learns the spatial relationships between different zones of the building. In block 2003, the time-series underlying model 1309 utilizes historical data from the building's monitored zones. The historical data from the monitored zones is the total energy usage over time corresponding to the monitored zone. In addition, in some embodiments, the historical data from the monitored zones includes the occupancy patterns of the monitored zones. In block 2005, the time-series underlying model 1309 extrapolates the historical data to infer disturbances in unmonitored zones.

[0116] Some embodiments are based on the understanding that, in addition to spatial relationships between different zones of a building, temporal correlations between different zones of a building can be used to improve disturbance predictions for unmonitored zones. For this reason, time-series-based models use spatial and temporal correlations between different zones of a building to identify interdependencies and improve disturbance predictions for unmonitored zones.

[0117] In some other embodiments, the time-series-based model 1309 is trained using a federated learning method. A federated learning method is a machine learning technique that allows a model to be trained across distributed devices or servers holding local data without sharing that data with a central server. Instead of collecting all the data in one place, a federated learning method trains the time-series-based model 1309 locally on individual devices (such as edge devices in multiple buildings or local servers), and only updates to the time-series-based model 1309 (i.e., weights or gradients) are shared with the central server. This helps maintain privacy and reduce the need for large-scale data transfers.

[0118] Examples of associative learning methods include horizontal associative learning, vertical associative learning, and associative transfer learning. In horizontal associative learning, data from multiple buildings has the same set of features but different sample sizes. In vertical associative learning, data from multiple buildings has different features but represents the same set of samples. Vertical associative learning allows for collaboration without sharing actual features. Associative transfer learning includes training models when data from multiple buildings are heterogeneous and inconsistent. Associative transfer learning enables learning by allowing learned knowledge about data types to be transferred to another.

[0119] In some further embodiments, the time-series-based model 1309 is trained using a multi-resolution learning technique. A multi-resolution learning technique is a method of training the time-series-based model 1309 on data of different levels of resolution, scale, or granularity. This strategy can improve the ability of the time-series-based model 1309 to capture both fine-grained details and broader patterns by providing different “views” or “perspectives” on the data. The multi-resolution technique allows the time-series-based model 1309 to learn from multiple levels of resolution or abstraction, ensuring that it can effectively generalize across data of varying scales.

[0120] For example, in one embodiment, building data is collected to build a model for predicting a building's energy consumption. The building data includes high-resolution data, i.e., hourly measurements such as energy consumption, temperature, and humidity. The building data also includes low-resolution data, i.e., daily aggregates such as energy consumption and average temperature. In a multi-resolution method, the model first learns from the low-resolution data (daily aggregates), which may capture broader trends such as seasonal patterns or weekly cycles. The model is then fine-tuned with the high-resolution data (hourly values), enabling it to capture more specific patterns such as daily peak consumption, sudden fluctuations, or local effects.

[0121] Some embodiments are based on the understanding that disturbances such as solar radiation, indoor environment occupancy patterns, and external temperature, which affect a building's energy consumption, are correlated. For example, solar radiation and external temperature are related; that is, strong solar radiation corresponds to high external temperature, and weak solar radiation corresponds to low external temperature. However, the time-series-based model 1309 does not capture the correlations between disturbances in order to predict them. As a result, the time-series-based model 1309 may inaccurately predict disturbances. For example, the time-series-based model 1309 may predict weak solar radiation and high external temperature.

[0122] Some embodiments are based on the understanding that the predictive accuracy of the time series base model 1309 can be improved by modifying it as a multivariate quantile function forecaster (MQF2). Figure 21 shows a feedback control system 1301 including an MQF2 2101 according to some embodiments of the present disclosure. In one embodiment, the MQF2 2101 is a module of the feedback control system 1301 and is executed by a processor 1305. The MQF2 2101 is an encoder-decoder method that enables a quantile-based representation of multivariate forecasts. The MQF2 2101 is trained to learn the correlations between disturbances as hidden vector outputs of the time series base model 1309, which is set up as an encoder. During estimation, the MQF2 2101 predicts disturbances by considering the correlations between disturbances based on the decoding of the hidden vector outputs from the time series base model 1309. In one embodiment, the hidden vector output from the time-series-based model 1309 indicates the disturbance predicted by the time-series-based model 1309.

[0123] Therefore, since disturbances are predicted considering the correlation between disturbances, MQF2 2101 predicts either (i) strong solar radiation and high external temperature, or (ii) weak solar radiation and low external temperature, as opposed to weak solar radiation and high external temperature. In other words, the accuracy of disturbance prediction is improved. Furthermore, based on such accurately predicted disturbances, the probabilistic feedback controller 1311 determines the control inputs to the GIB system 1303 that optimize the performance of the GIB system 1303 while achieving the desired indoor environmental conditions. Since the probabilistic feedback controller 1311 determines the control inputs based on such accurately predicted disturbances, the performance of the GIB system 1303 is effectively optimized while precisely achieving the desired indoor environmental conditions.

[0124] In some embodiments, the time-series-based model 1309 is a Temporal Fusion Transformer (TFT) model, and the TFT model is modified to form an MQF2. For example, a partially input convex neural network (PICNN) is added to the TFT model. Furthermore, the input / output structure of the TFT model is modified. In some embodiments, modifying the output structure of the TFT model involves changing the output vector of the TFT model at each time t+τ to a hidden output vector h t+τ This includes remaking it as such.

[0125] Figure 22A shows a block diagram of a TFT2200 forming an MQF2 2101 according to some embodiments of this disclosure. The TFT2200 includes a static covariate encoder 2201, a variable selection network (VSN) 2203, a gated residual network 2205, and a time-fusion decoder 2207. Each of these will be described below.

[0126] Static Covariate Encoder 2201: The static covariate encoder 2201 is configured to integrate static features into an architecture through the encoding of context vectors to modulate temporal dynamics. To handle multivariate cases, i.e., N target scalar variables represented by target vectors y∈R instead of a single scalar variable, several modifications to the original TFT model are required. The VSN, which aggregates all target and covariate inputs independently for each timestamp, can be easily adapted to accept target vectors by adding the required input channels to each VSN. More specifically, this simply means that during past time tj=tk…t, the encoder VSN input x t-j In the target scalar y t-j The target vector y t-j This means that it will be replaced by.

[0127] The VSN2203 is configured to select relevant input variables at each time step. The gated residual network 2205 is configured to skip any unused components of the architecture, providing adaptive depth and network complexity to accommodate a wide range of datasets and scenarios.

[0128]

number

[0129]

number

[0130] The objective of some embodiments is to consider the correlation between disturbances predicted at different times, i.e., to model temporal correlations.

[0131] Some embodiments are based on the understanding that correlations between disturbances predicted at different times can be taken into account by modifying the input to the time-series-based model 1309.

[0132] Figure 22B shows a modified input 2210 to a time-series-based model 1309 according to some embodiments of the present disclosure. The modified input 2210 includes a past disturbance forecast 2209 and a current input 2211. The past disturbance forecast 2209 corresponds to disturbances predicted at past points in time. The past disturbance forecast 2209 may be stored in memory 1307 or retrieved from external memory. The current input 2211 includes real-time data 1315. The past disturbance forecast 2209 and the current input 2211 are concatenated (2213) to form the modified input 2210 to the time-series-based model 1309.

[0133] The modified input 2210 is input to the time-series underlying model 1309. Based on the modified input 2210, the time-series underlying model 1309 predicts new disturbances affecting the building's energy consumption. Predicting new disturbances based on past disturbance predictions 2209 and the current input 2211 makes the disturbance predictions temporally autoregressive.

[0134]

number

[0135]

number

[0136]

number

[0137] Figure 23 is a schematic diagram showing a computing device for implementing the method and system of this disclosure as a non-limiting example. The computing device 2300 may include a power supply 2301, a processor 2303, a memory 2305, and a storage device 2307, all connected to a bus 2309. Furthermore, a high-speed interface 2311, a low-speed interface 2313, a high-speed expansion port 2315, and a low-speed connection port 2317 may be connected to the bus 2309. In addition, a low-speed expansion port 2319 is connected to the bus 2309. Furthermore, an input interface 2321 may be connected to an external receiver 2323 and an output interface 2325 via the bus 2309. Receiver 2327 may be connected to external transmitters 2329 and 2331 via the bus 2309. External memory 2333, an external sensor 2335, a machine 2337, and an environment 2339 may also be connected to the bus 2309. Furthermore, one or more external input / output devices 2341 can be connected to bus 2309. A network interface controller (NIC) 2343 can be adapted to connect to network 2345 via bus 2309, and in particular, data or other data can be rendered on third-party display devices, third-party imaging devices, and / or third-party printing devices located outside the computer device 2300.

[0138] Memory 2305 can store instructions that the computer device 2300 can execute, historical data, and any data available to the methods and systems of this disclosure. Memory 2305 may include random access memory (RAM), read-only memory (ROM), flash memory, or any other suitable memory system. Memory 2305 may consist of one or more volatile memory units and / or one or more non-volatile memory units. Memory 2305 may also consist of another form of computer-readable medium, such as a magnetic disk or an optical disk.

[0139] The storage device 2307 can be adapted to store supplemental data and / or software modules used by the computer device 2300. For example, the storage device 2307 can store historical data and other relevant data previously described in relation to this disclosure. In addition to or instead of that, the storage device 2307 can store historical data such as the data previously described in relation to this disclosure. The storage device 2307 may include a hard drive, an optical drive, a thumb drive, an array of drives, or any combination thereof. Furthermore, the storage device 2307 may include computer-readable media such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, flash memory or other similar solid-state memory device, or an array of devices including a storage area network or other configuration. Instructions can be stored in an information carrier. When an instruction is executed by one or more processing devices (e.g., processor 2303), it performs one or more of the methods described above.

[0140] The computing device 2300 may optionally be linked via bus 2309 to a display interface or user interface (HMI) 2347 adapted to connect the computing device 2300 to a display device 2349 and a keyboard 2351, the display device 2349 may include, among other things, a computer monitor, a camera, a television, a projector, or a mobile device. In some implementations, the computer device 2300 may also include a printer interface for connecting to a printing device, the printing device may include, among other things, a liquid inkjet printer, a solid ink printer, a large-scale commercial printer, a thermal printer, a UV printer, or a dye-sublimation printer.

[0141] The high-speed interface 2311 manages the bandwidth-intensive operation of the computing device 2300, and the low-speed interface 2313 manages the low-bandwidth-intensive operation. Such function assignments are merely examples. In some implementations, the high-speed interface 2311 can be coupled to memory 2305, user interface (HMI) 2347, keyboard 2351, and display 2349 (e.g., through a graphics processor or accelerator), and can also be coupled to a high-speed expansion port 2315 that can accept various expansion cards via bus 2309. In some implementations, the low-speed interface 2313 is coupled to storage device 2307 and low-speed expansion port 2319 via bus 2309. The low-speed expansion port 2319, which may include various communication ports (e.g., USB, Bluetooth®, Ethernet®, Wireless Ethernet®), may be connected to one or more input / output devices 2341. The computing device 2300 may be coupled to server 2353 and rack server 2355. The computing device 2300 may be implemented in several different forms. For example, the computing device 2300 may be implemented as part of a rack server 2355.

[0142] This specification provides only specific embodiments and is not intended to limit the scope, applicability, or configuration of the disclosure. Rather, the following description of specific embodiments will provide a description that enables the realization of one or more specific embodiments for those skilled in the art. Various modifications are intended to be made to the function and configuration of the elements without departing from the spirit and scope of the subject matter disclosed in the appended claims.

[0143] Specific details are provided in the following description to ensure a full understanding of the embodiments. However, those skilled in the art will understand that the embodiments can be carried out even without these specific details. For example, systems, processes, and other elements in the disclosed subject matter may be shown as components in the form of block diagrams to avoid obscuring the embodiments with unnecessary details. In other examples, well-known processes, structures, and techniques may be shown without unnecessary details to avoid obscuring the embodiments. Furthermore, similar reference numbers and names in different drawings refer to similar elements.

[0144] Furthermore, individual embodiments may be described as processes shown as flowcharts, flow diagrams, data flow diagrams, structural diagrams, or block diagrams. While flowcharts may describe operations as sequential processes, many operations can be performed in parallel or simultaneously. In addition, the order of operations may be rearranged. A process may terminate when its operations are complete, but it may have additional steps that are not discussed or included in the diagrams. Moreover, not all operations in any process specifically described can occur in all embodiments. A process may correspond to a method, function, procedure, subroutine, subprogram, etc. If a process corresponds to a function, the termination of the function may correspond to returning the function to a calling function or main function.

[0145] Furthermore, embodiments of the disclosed subject matter may be implemented either manually or automatically, at least in part. Manual or automatic implementation may be performed, or at least assisted, through a machine, hardware, software, firmware, middleware, microcode, hardware description language, or any combination thereof. If implemented in software, firmware, middleware, or microcode, the program code or code segments for performing the required tasks may be stored in machine-readable media. A processor(s) may perform the required tasks.

[0146] The various methods or processes outlined herein may be encoded as software executable on one or more processors employing any one of a variety of operating systems or platforms. In addition, such software may be written using any of several suitable programming languages ​​and / or programming or scripting tools, and may be compiled as executable machine language code or intermediate code that runs on a framework or virtual machine. Typically, the functions of program modules may be combined or distributed as desired in various embodiments.

[0147] Embodiments of this disclosure may be implemented as methods, and an example thereof is provided. The order of operations performed as part of this method may be determined in any suitable manner. Thus, embodiments may be configured such that operations are performed in an order different from the order illustrated, which may include performing some operations simultaneously, although they are shown as a series of operations in the illustrated embodiments.

[0148] Furthermore, embodiments of the present disclosure and the functional operations described herein can be implemented in digital electronic circuits, in tangibly implemented computer software or firmware, in computer hardware including the structures disclosed herein and their structural equivalents, or in one or more combinations thereof. Furthermore, some embodiments of the present disclosure can be implemented as one or more computer programs, i.e., as one or more modules of computer program instructions encoded on a tangible, non-temporary program carrier for execution by a data processing device or for controlling the performance of a data processing device. Furthermore, the program instructions can be encoded on artificially generated propagating signals, for example, on machine-generated electrical, optical, or electromagnetic signals. The propagating signals are generated to encode information to be transmitted to a suitable receiving device for execution by the data processing device. The computer storage medium may be a machine-readable storage device, a machine-readable storage board, a random-access memory device, or a serial-access memory device, or one or more combinations thereof.

[0149] In accordance with embodiments of this disclosure, the term “data processing device” may encompass all types of devices, machines, and apparatus that process data, including, for example, a programmable processor, a computer, or multiple processors or computers. The device may include dedicated logic circuits, such as FPGAs (field-programmable gate arrays) or ASICs (application-specific integrated circuits). In addition to hardware, the device may also include code that generates the execution environment for the computer program, such as processor firmware, a protocol stack, a database management system, an operating system, or code comprising one or more of these.

[0150] Computer programs (sometimes called or described as programs, software, software applications, modules, software modules, scripts, or code) can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages, and can be deployed in any form, as standalone programs or as modules, components, subroutines, or other units suitable for use in a computing environment. Computer programs may or may not correspond to files in a file system. A program may be stored in part of a file that holds other programs or data, for example, one or more scripts stored in a markup language document, a single file dedicated to the program in question, or a coordinated set of files, for example, one or more modules, subprograms, or parts of code.

[0151] Computer programs can be deployed to run on a single computer, or on multiple computers located in one place or distributed across multiple locations and interconnected by a communication network. A computer suitable for running a computer program may, for example, be based on a general-purpose microprocessor, a dedicated microprocessor, or both, or any other type of central processing unit. Generally, the central processing unit receives instructions and data from read-only memory, random-access memory, or both. Essential elements of a computer are a central processing unit for executing or running instructions, and one or more memory devices for storing instructions and data.

[0152] Generally, a computer also includes one or more mass storage devices for storing data, such as magnetic, magneto-optical, or optical disks, or is operationally coupled to such disks to receive data from them, transfer data to them, or both. However, a computer does not have to have such devices. Furthermore, a computer can be embedded in another device. For example, it can be embedded in a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device, such as a Universal Serial Bus (USB) flash drive.

[0153] To provide user interaction, embodiments of the subject matter described herein may be implemented on a computer having a display device for displaying information to the user, such as a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, and a keyboard and pointing device, such as a mouse or trackball, that allows the user to provide input to the computer. User interaction may be provided using other types of devices. For example, the feedback provided to the user may be any form of sensory feedback, such as visual feedback, auditory feedback, or tactile feedback, and input from the user may be received in any form, including acoustic input, voice input, or tactile input. In addition, the computer may implement user interaction by sending documents to and receiving documents from a device used by the user, for example, by sending web pages to a web browser on the user's client device in response to a request received from the web browser.

[0154] Embodiments of the subject matter described herein can be implemented in a computing system that includes, for example, a backend component as a data server, or a middleware component, such as an application server, or a frontend component, such as a client computer having a graphical user interface or a web browser that allows a user to interact with an implementation of the subject matter described herein, or any combination of one or more such backend, middleware, or frontend components. The components of the system can be interconnected by any form or medium of digital data communication, such as a communication network. Examples of communication networks include local area networks ("LANs") and wide area networks ("WANs"), such as the Internet.

[0155] A computing system can include clients and servers. Clients and servers are generally geographically separated and typically communicate through a communication network. The client-server relationship arises from computer programs running on each computer that have a client-server relationship.

[0156] While this disclosure has been described using several preferred embodiments, it should be understood that various other adaptations and modifications can be carried out within the spirit and scope of this disclosure. Therefore, the following claims are intended to cover all such variations and modifications that fall within the true spirit and scope of this disclosure.

Claims

1. A feedback control system for optimizing the performance of a grid-interactive building (GIB) system configured to regulate the indoor environment of a building, wherein the feedback control system is: A time-series underlying model configured to predict disturbances affecting the energy consumption of the aforementioned building, The system comprises a probabilistic feedback controller, which is configured to determine a control input to the GIB system by evaluating a plurality of control actions to the GIB system based on the disturbances affecting the predicted energy consumption of the building, the control input maximizing the likelihood of achieving desired indoor environmental conditions while minimizing energy consumption, and the feedback control system further comprises It is equipped with a feedback mechanism, and the feedback mechanism is We collect real-time data including indoor environmental conditions, actual energy consumption, and external factors. The time-series underlying model is fine-tuned using the collected real-time data. A feedback control system configured to update the probabilistic feedback controller based on the difference between predicted data and the collected real-time data.

2. The feedback control system according to claim 1, wherein the probabilistic feedback controller evaluates the plurality of control actions on the GIB system using Bayesian estimation.

3. The feedback control system according to claim 1, wherein the probabilistic feedback controller evaluates the plurality of control actions on the GIB system using probabilistic optimization techniques.

4. The feedback control system according to claim 1, wherein the time-series-based model is pre-trained on a general-purpose dataset.

5. The feedback control system according to claim 4, wherein the time-series-based model, pre-trained on the general-purpose dataset, is fine-tuned using building-specific historical data to fit the building's energy consumption patterns.

6. The feedback control system according to claim 1, wherein the feedback mechanism is further configured to update the time-series underlying model based on the real-time data using a reinforcement learning framework.

7. The feedback control system according to claim 1, wherein the probabilistic feedback controller and the time-series-based model are jointly optimized to minimize energy consumption while maintaining indoor environmental conditions within a predetermined range.

8. The feedback control system according to claim 1, wherein the time series base model is fine-tuned using low-rank approximation to reduce the complexity of the time series base model while retaining essential predictive features.

9. The fine-tuning using the aforementioned low-rank approximation is, Identifying low-rank subspaces within the feature representation of the aforementioned time-series-based model, Updating the low-rank subspace during fine-tuning, The feedback control system according to claim 8, comprising maintaining fixed, pre-trained, high-rank parameters to preserve general knowledge of the time-series underlying model.

10. The feedback control system according to claim 8, wherein the low-rank approximation is applied to a specific layer of the time-series base model to fit the time-series base model to the building's energy consumption pattern while preserving a pre-trained representation.

11. The feedback control system according to claim 8, wherein the low-rank approximation is dynamically adjusted during the fine-tuning based on the available computing resources of the feedback control system or the size of the building-specific historical data.

12. By learning global patterns from data collected from multiple buildings, Fine-tune building-specific historical data to match the building's energy consumption patterns. The feedback control system according to claim 1, wherein the time-series-based model is trained hierarchically.

13. The feedback control system according to claim 1, wherein the time-series-based model is fine-tuned using data from other buildings having similar structural or operational characteristics to predict the disturbance.

14. The aforementioned time-series-based model detects disturbances in the unmonitored zones of the building, Learn the spatial relationships between different zones in the aforementioned building, Utilizing historical data from the monitored zones of the aforementioned building, The aforementioned historical data is extrapolated to estimate disturbances in the unmonitored zone. The feedback control system according to claim 1, which predicts by doing so.

15. The feedback control system according to claim 14, wherein the time-series-based model uses the spatial and temporal correlations between the different zones of the building to identify interdependencies and improve prediction of disturbances for the unmonitored zones.

16. The feedback control system according to claim 1, wherein the time-series-based model is trained using one of a federated learning method and a multi-resolution learning method.

17. The feedback control system according to claim 16, wherein the time-series-based model in the multi-resolution learning method is trained with data of different resolutions and scales.

18. The feedback control system according to claim 1, further comprising a multivariate quantile predictor, wherein the multivariate quantile predictor is configured to predict disturbances affecting the energy consumption of the building by considering the correlations between the disturbances based on the hidden vector outputs of the time series underlying model.

19. The feedback control system according to claim 18, wherein the multivariate quantile predictor is a modified time-fusion transformer (TFT) model.

20. The feedback control system according to claim 1, wherein the time-series-based model is further configured to predict new disturbances that affect the energy consumption of the building in a time-regressive manner, based on previously predicted disturbances and the real-time data.

21. A method for optimizing the performance of a grid-interactive building (GIB) system configured to regulate the indoor environment of a building, wherein the method is: The time-series underlying model predicts disturbances that affect the energy consumption of the building, The probabilistic feedback controller includes determining a control input to the GIB system by evaluating a plurality of control actions to the GIB system based on the disturbances affecting the predicted energy consumption of the building, wherein the control input maximizes the likelihood of achieving desired indoor environmental conditions while minimizing energy consumption, and the method further includes: The feedback mechanism collects real-time data including indoor environmental conditions, actual energy consumption, and external factors. The feedback mechanism uses the collected real-time data to fine-tune the time-series-based model, A method comprising the feedback mechanism updating the probabilistic feedback controller based on the difference between predicted data and the collected real-time data.

22. A non-temporary computer-readable storage medium incorporating a processor-executable program for performing a method for optimizing the performance of a grid-interactive building (GIB) system configured to regulate the indoor environment of a building, wherein the method is: The time-series underlying model predicts disturbances that affect the energy consumption of the building, The probabilistic feedback controller includes determining a control input to the GIB system by evaluating a plurality of control actions to the GIB system based on the disturbances affecting the predicted energy consumption of the building, wherein the control input maximizes the likelihood of achieving desired indoor environmental conditions while minimizing energy consumption, and the method further includes: The feedback mechanism collects real-time data including indoor environmental conditions, actual energy consumption, and external factors. The feedback mechanism uses the collected real-time data to fine-tune the time-series-based model, A non-temporary computer-readable storage medium, wherein the feedback mechanism updates the probabilistic feedback controller based on the difference between predicted data and the collected real-time data.

Images