Abstract: Systems and methods are provided for efficiently building an object detection learning model for an unlabeled pool of images. A recommendation engine automatically recommends an annotation type for the images in the unlabeled pool based on previous object detection and an updated mean average precision of the model, where the mean average precision represents the performance of the model.

Abstract: A method of generating a fault determination in a building management system (BMS), the method including receiving signal data, generating, using a number of fault detection models, a number of fault indications based on the signal data, generating, using a weighting function, based on the number of fault indications, a fault score, comparing the fault score to a fault value, and determining, based on the comparison, an existence of a fault.

Abstract: A model management system for building equipment includes one or more memory devices configured to store instructions that, when executed on one or more processors, cause the one or more processors to determine whether fault data exists in equipment data used to generate a plurality of shutdown prediction models for the building equipment, generate a first performance evaluation value for each of the plurality of shutdown prediction models using a first evaluation technique in response to a determination that the fault data exists in the equipment data, generate a second performance evaluation value for each of the plurality of shutdown prediction models using a second evaluation technique in response to a determination that the fault data does not exist in the equipment data, and select one of the plurality of shutdown prediction models based on the first performance evaluation value and the second performance evaluation value.

Abstract: A method of operating a building management system is disclosed. The method includes determining, by a processing circuit, policy rankings for a plurality of control policies based on building operation data of a first previous time period, selecting, by the processing circuit, a set of control policies from among the plurality of control policies based on the policy rankings of the set of control policies satisfying a ranking threshold, generating, by the processing circuit, a plurality of prediction models for the set of control policies, selecting, by the processing circuit, a first prediction model of the plurality of prediction models based on building operation data of a second previous time period, and responsive to selecting the first prediction model, operating, by the processing circuit, the building management system using the first prediction model.

Abstract: A building system operates to receive building data for a building describing one or more conditions of the building and perform a first optimization with a multi-tiered model that predicts a first condition of the building based on a first control setting, the first optimization determining one or more first values of the first control setting. The building system operates to perform a second optimization with the multi-tiered model that predicts a second condition of the building based on a second control setting and the one or more first values of the first control setting, the second optimization determining one or more second values of the second control setting and operate building equipment based on the one or more first values of the first control setting and the one or more second values of the second control setting.

Abstract: Systems and methods for training a reinforcement learning (RL) model for HVAC control are disclosed herein. Simulated experience data for the HVAC system is generated or received. The simulated experience data is used to initially train the RL model for HVAC control. The HVAC system operates within a building using the RL model and generates real experience data. A determination may be made to retrain the RL model. The real experience data is used to retrain the RL model. In some embodiments, both the simulated and real experience data are used to retrain the RL model. Experience data may be sampled according to various sampling functions. The RL model may be retrained multiple times over time. The RL model may be retrained less frequently over time as more real experience data is used to train the RL model.

Abstract: A model management system for building equipment includes one or more memory devices configured to store instructions that, when executed on one or more processors, cause the one or more processors to determine whether fault data exists in equipment data used to generate a plurality of shutdown prediction models for the building equipment, generate a first performance evaluation value for each of the plurality of shutdown prediction models using a first evaluation technique in response to a determination that the fault data exists in the equipment data, generate a second performance evaluation value for each of the plurality of shutdown prediction models using a second evaluation technique in response to a determination that the fault data does not exist in the equipment data, and select one of the plurality of shutdown prediction models based on the first performance evaluation value and the second performance evaluation value.

Abstract: A building system for training a prediction model with augmented training data. The building system comprising one or more memory devices configured to store instructions thereon that, when executed by one or more processors, cause the one or more processors to obtain a first training data set comprising data values associated with a data point of the building system and with a plurality of time-steps and energy values associated with consumption of the building system at each of the plurality of time-steps; generate an augmented training data set comprising a second training data set, the second training data set comprising the energy values and the data values of the first training data set but with a data value replaced with a predetermined value at a time-step of the plurality of time-steps; and generate a prediction model by training the prediction model.

Abstract: A fault prediction system for building equipment includes one or more memory devices configured to store instructions that, when executed on one or more processors, cause the one or more processors to receive device data for a plurality of devices of the building equipment, the device data indicating performance of the plurality of devices; generate, based on the received device data, a plurality of prediction models comprising at least one of single device prediction models generated for each of the plurality of devices or cluster prediction models generated for device clusters of the plurality of devices; label each of the plurality of prediction models as an accurately predicting model or an inaccurately predicting model based on a performance of each of the plurality of prediction models; and predict a device fault with each of the plurality of prediction models labeled as an accurately predicting model.

Abstract: A building system for detecting faults in an operation of building equipment. The building system comprising one or more memory devices configured to store instructions thereon that cause the one or more processors to perform a cumulative sum (CUSUM) analysis on actual building data and corresponding predicted building data to obtain cumulative sum values for a first plurality of times within a first time period; analyze cumulative sum values associated with a second plurality of times occurring before the first time to identify a second time of the second plurality of times at which a second cumulative sum value is at a local minimum; and determine that a first fault began at the second time.

Abstract: A building system for generating input forecast data. The building system comprising one or more memory devices configured to store instructions thereon that, when executed by one or more processors, cause the one or more processors to retrieve a current prediction set of measurements comprising current values associated with a plurality of time-steps; identify time-steps of the plurality of time-steps for which historical prediction values of a historical prediction set of measurements are outside of a tolerance of corresponding historical actual values of a historical actual set of measurements; replace each current value of the current prediction set of measurements that is associated with the identified time-steps with a predetermined value to generate an updated prediction set of measurements; and provide the updated prediction set of measurements to a prediction model.

Abstract: A chiller threshold management system for a building, including one or more memory devices and one or more processors. The one or more memory devices are configured to store instructions to be executed on the one or more processors. The one or more processors are configured to determine whether chiller fault data exists in chiller data used to generate a plurality of chiller prediction models. The one or more processors are further configured to generate a first threshold evaluation value for each of the plurality of chiller prediction models using a first evaluation technique in response to a determination that chiller fault data exists in the chiller data, and generate a second threshold evaluation value for each of the chiller prediction models using a second evaluation technique in response to a determination that chiller fault data does not exist in the chiller data.

Abstract: Systems and methods for training a reinforcement learning (RL) model for HVAC control are disclosed herein. A calibrated simulation model is used to train a surrogate model of the HVAC system operating within a building. The surrogate model is used to generate simulated experience data for the HVAC system. The simulated experience data can be used to train a reinforcement learning (RL) model of the HVAC system. The RL model is used to control the HVAC system based on the current state of the system and the best predicted action to perform in the current state. The HVAC system generates real experience data based on the actual operation of the HVAC system within the building. The real experience data is used to retrain the surrogate model, and additional simulated experience data is generated using the surrogate model. The RL model can be retrained using the additional simulated experience data.

Abstract: A fault prediction system for building equipment includes one or more memory devices configured to store instructions that, when executed on one or more processors, cause the one or more processors to receive device data for a plurality of devices of the building equipment, the device data indicating performance of the plurality of devices; generate, based on the received device data, a plurality of prediction models comprising at least one of single device prediction models generated for each of the plurality of devices or cluster prediction models generated for device clusters of the plurality of devices; label each of the plurality of prediction models as an accurately predicting model or an inaccurately predicting model based on a performance of each of the plurality of prediction models; and predict a device fault with each of the plurality of prediction models labeled as an accurately predicting model.

Abstract: Systems and methods for training a reinforcement learning (RL) model for HVAC control are disclosed herein. Simulated experience data for the HVAC system is generated or received. The simulated experience data is used to initially train the RL model for HVAC control. The HVAC system operates within a building using the RL model and generates real experience data. A determination may be made to retrain the RL model. The real experience data is used to retrain the RL model. In some embodiments, both the simulated and real experience data are used to retrain the RL model. Experience data may be sampled according to various sampling functions. The RL model may be retrained multiple times over time. The RL model may be retrained less frequently over time as more real experience data is used to train the RL model.

Abstract: A model management system for a building, including one or more memory devices and one or more processors. The one or more memory devices are configured to store instructions to be executed on the one or more processors. The one or more processors are configured to determine whether chiller fault data exists in chiller data used to generate a plurality of chiller shutdown prediction models. The one or more processors are further configured to generate a first performance evaluation value for each of the plurality of chiller shutdown prediction models using a first evaluation technique in response to a determination that chiller fault data exists in the chiller data, and generate a second performance evaluation value for each of the plurality of chiller shutdown prediction models using a second evaluation technique in response to a determination that chiller fault data does not exist in the chiller data.

Abstract: A building system for building data point prediction, the building system comprising one or more memory devices configured to store instructions, that, when executed by one or more processors, cause the one or more processors to receive first building data for a building data point of a building and generate training data, the training data comprising a probability distribution sequence comprising a first probability distribution for the building data point. The instructions cause the one or more processors to train a prediction model based on the training data, receive second building data for the building data point, and predict, for one or more time-steps into the future, one or more second probability distributions with the second building data based on the prediction model, each of the one or more second probability distributions being a probability distribution for the building data point at one of the one or more time-steps.

Abstract: Systems and methods for training a reinforcement learning (RL) model for HVAC control are disclosed herein. Simulated experience data for the HVAC system is generated or received. The simulated experience data is used to initially train the RL model for HVAC control. The HVAC system operates within a building using the RL model and generates real experience data. A determination may be made to retrain the RL model. The real experience data is used to retrain the RL model. In some embodiments, both the simulated and real experience data are used to retrain the RL model. Experience data may be sampled according to various sampling functions. The RL model may be retrained multiple times over time. The RL model may be retrained less frequently over time as more real experience data is used to train the RL model.

Abstract: A chiller fault prediction system for a building, including one or more memory devices and one or more processors. The one or more memory devices are configured to store instructions to be executed on the one or more processors. The one or more processors are configured to receive chiller data for a plurality of chillers, the chiller data indicating performance of the plurality of chillers. The one or more processors are configured to generate, based on the received chiller data, a plurality of single chiller prediction models and a plurality of cluster chiller prediction models, the plurality of single chiller prediction models generated for each the plurality of chillers and the plurality of cluster chiller prediction models generated for chiller clusters of the plurality of chillers.

Abstract: Methods, systems, and computer storage media are disclosed for determining electric energy flow predictions for electric systems including photovoltaic solar systems. In some examples, a method is performed by a computer system and includes supplying a consumption time series and a predicted production time series for an electric system to a machine-learning predictor trained during a prior training phase using electric energy consumption training data and photovoltaic production training data. The consumption time series has a first data resolution, and the electric energy consumption training data and the photovoltaic production training data have a second data resolution greater than the first data resolution. The method includes determining, using an output of the machine-learning predictor, a predicted import time series of electric import values each specifying an amount of electric energy predicted to be imported by the electric system with a prospective photovoltaic solar system installed.