Modeling of an Autonomous AI HVAC Control System Commissioning
Hi everyone! Brain4Energy is live!
About us
We are a team of data scientists, who develop autonomous building HVAC (Heating, Ventilation, and Air Conditioning) control systems. Our AI-based adaptive control policies allow the building owners to decrease building energy consumption.
Research Issue
In this article, we want to share our modeling results of the commissioning of such an AI HVAC control system. More precisely, we want to review a historical data insufficiency issue, which appears due to the simple control policy of a default building HVAC control system by PID controller supervision, e.g. target zone temperature may be 22℃ for weekdays and 26℃ for the weekend for the whole year. And of course, this data is not enough to learn the HVAC equipment behavior, so we need more various data.
Simulation Background
We will conduct the issue research on the West Zone of the Two-zone Data Center model with the EnergyPlus simulation environment. The East Zone always will be under constant control. The scheme of the Two-zone Data Center model and interaction with the model are shown in Fig. 1a and Fig. 1b.
The necessary HVAC equipment of the West Zone includes the next items:
- outdoor air system (OA System) exchanges zone and outdoor airflow;
- variable volume fan (VAV Fan) adjusts the airflow rate to meet the Zone Mean Air temperature;
- a direct evaporative cooler (DEC) lowers the air temperature using latent heat from the evaporation of water;
- an indirect evaporative cooler (IEC) lowers the air temperature without adding humidity to the air;
- direct expansion cooling coil (DX CC) cools air passing the condensed refrigerant through a heat exchanger;
- chilled water cooling coil (CW CC) cools the air using chilled water (not used in the West Zone);
- zone thermostat (ZT) provides Zone Mean Air temperature.
Applying Model Predictive Control
Research works about controlling the HVAC systems are aimed to decrease HVAC equipment energy consumption using different control methods. The most promising are based on reinforcement learning and model predictive control (MPC) techniques. In this article, we will use the MPC approach (Fig. 2).
According to this technic, the MPC controller at each moment of making a control decision (timestep) predicts a list of future HVAC equipment sensors states (e.g. temperature, humidity, etc.) and solves the optimization task required by the objective (decrease energy consumption) and constraints based on predicted states (to remain in desired temperature range).
As a state predictive model of the MPC controller, we will apply the multitask perceptron model. with the architecture shown in Fig. 3. Each hidden layer contains an attention layer with a parametric ReLU activation function. The predicted states are:
- Zone Mean Air Temperature;
- ITE CPU Electric Power;
- HVAC Power Consumption.
The list of states and setpoints for training and predictions with current states and history depth of 20 timesteps states for each item is:
- Site Outdoor Air Dry Bulb Temperature;
- Zone Mean Air Temperature;
- ITE CPU Electric Power;
- HVAC Power Consumption;
- System Node Setpoint Temperature;
- System Node Setpoint Temperature Difference;
- Fan Air Mass Flow Rate;
- Fan Air Mass Flow Rate Difference.
The background part is over, let’s try to save some energy!
Insufficient Data Issue
Firstly, we will take historical data for a month along with June Chicago weather data and will try to decrease energy consumption. The results of the comparison between the PID controller and the multitask perceptron are shown in Fig. 4. The PID controller is adjusted to keep a higher position in the target temperature range to consume less power. Despite the achieved Power Consumption (PC) mean daily economy of 11.49% of the multitask perceptron, the desired Zone Mean Air Temperature range (green bar: 23±1℃) was significantly violating (44 % of out of the range points). And the reason is poor training dataset.
Insufficient Data Issue Solution
To collect more various data about HVAC equipment behavior, we were arbitrary “twitching” HVAC equipment setpoints while keeping Zone Mean Air Temperature in the desired range. For this purpose, we used machine learning random walk methods with predictive models trained on the default PID control’s historical data. Then, with the collected enhanced dataset, we trained the MPC controller and assessed the control performance.
Simulation Design
Experimental Workflow
The experimental workflow is shown in Fig. 5. To “twitch” HVAC equipment setpoints in a safe manner, we need a model that can predict the HVAC equipment sensors states and provide the desired Zone Mean Air Temperature range. We used a random walk method to generate control value and 3 different models to predict the Zone Mean Air Temperature possible value. These models are not able to provide much accuracy, but they are accurate enough to evaluate if the temperature is in a safe range or not. Also, they are simple and easy to train. Besides, Bayesian Neural Net and Gaussian Process Regressor provide probability distribution (possible temperature range). The influence of these 3 model types will be evaluated in the MPC performance assessment. Here’s the models summary:
- Feedforward neural network from the PyTorch library with parameters:
- Input size: 16 (history depth — 2 steps);
- 2 convolution layers;
- loss function: logcosh;
- learning rate: 0.5;
- ‘batch_size’: 16;
- ‘epoch’: 150;
- optimizer: Adadelta (pytorch).
Bayesian Neural Net with parameters:
- history depth: 6;
- loss function: mean squared error;
- metrics: MSELoss (pytorch);
- optimizer: Adam (pytorch);
- activation function: PReLU (pytorch);
Gaussian process regressor from the scikit-learn library with parameters:
- history depth: 10;
- kernel: Exp-Sine-Squared (scikit-learn) with length scale = 1,
periodicity = 5, and periodicity_bounds (0.01; 10) with white noise
(noise level = 0.1).
For training to predict Zone Mean Air Temperature for the random walk methods we used the next states and setpoints as the models’ features:
- Site Outdoor Air;
- Zone Mean Air Temperature;
- ITE CPU Electric Power;
- HVAC Power Consumption;
- Air Loop Outlet Node Temperature Setpoint;
- Air Loop Outlet Node Temperature Difference;
- Zone Supply Fan Air Mass Flow Rate Setpoint;
- Zone Supply Fan Air Mass Flow Rate Difference.
After the training stage, each random walk method was run on 31 days of data to generate the enhanced dataset. The algorithm of setpoints selection is given in Fig. 6.
Each random walk model predicts possible Zone Mean Air Temperature range after applying randomly changed setpoints. If the range is in desired limits, a random walk algorithm uses these setpoints, otherwise, it repeats the iteration.
Temperature setpoints were randomly changing from the current setpoints’ values on a value up to ±3 ℃ (discrete is 0.5 ℃) in the range [10 ℃; 19 ℃].
Airflow setpoints were randomly changing from the current setpoints’ values on a value up to ±1.6 kg/s (discrete is 0.2 kg/s) in range [3 kg/s; 6.2 kg/s].
Then, using the enhanced dataset, we trained the MPC controller and assessed its performance a month after.
Evaluation Metrics
Metrics to evaluate random walk generated data:
- Zone Mean Air temperature range violation coefficient: relation of the out of the desired range points number to all points number;
- Mean daily economy coefficient.
For the mean daily economy coefficient calculations, the PID control power consumption is taken as the baseline.
Results
The resulting plot of applying random walk methods are shown in Fig. 7–9. The evaluation metrics for each plot are given in the plots’ titles.
Comparing the result of the multitask perceptron learned on historical data generated by PID control (Fig. 4) and the results of the multitask perceptron learned on the enhanced datasets (Fig. 7–9), it seen that we have improved the violation coefficient from 44% to 33% keeping energy saving values at the same level about 11.5%.
The results of the linear perceptron (Fig. 7) has more violations of the target temperature zone (~37%) than the BNN (~33%) and the GP model (~34%), but the linear perceptron has the easiest configuring, the fastest processing and has shown a bit greater power efficiency (~0.5%).
The results of the BNN and the GP model look similar, but the BNN performs more stable, which can be visually seen in Fig. 8.
Unlike the PID controller, all applied models have violated the target temperature zone. This is typical for HVAC AI control systems based on machine learning algorithms (e.g. Building HVAC Scheduling Using Reinforcement Learning via Neural Network Based Model Approximation) and presents itself as a trade-off between power efficiency and occupants comfort.
This paper describes our first research iteration of HVAC AI control system commissioning without any additional approaches. It is seen that each random walk method has improved MPC control quality keeping energy saving values at the same level. Considering that the enhanced dataset generation period lasted 31 days, the achieved results can be considered well-promising.
As our experience of such HVAC AI control systems commissioning in real buildings has shown, building equipment has more potential for power consumption decreasing. The real daily economy numbers can be achieved by up to 40%.
Future works
In the upcoming works, we are going to describe a real case of Autonomous AI HVAC Control System Commissioning and a case of modeling with a more sophisticated model.
