Technical Article

# Using AI/ML To Predict Brownouts in Step-Down Transformers

April 07, 2024 by Dr. Don Wilcher

## This article explores how artificial intelligence and machine learning predict brownout thresholds in a step-down transformer using a circuit simulator to obtain data and build an ML classifier.

Maintaining a stable power grid requires anticipating potential issues before they disrupt operations. Step-down transformers, crucial for delivering usable voltage levels to consumers, are susceptible to brownout conditions—a decrease in voltage that can affect equipment performance.

### The Challenge: Identifying Brownout Conditions

The laboratory exercise focused on a step-down transformer with a primary winding supply voltage of 120 Vac. The primary voltage was systematically reduced in stages: initially in 5-volt decrements and then in 10-volt decrements after reaching 100 Vac. The secondary output voltage was adjusted based on the transformer's 5:1 turns ratio. Developing a model to predict the lowest voltage at which the transformer could operate normally before experiencing a brownout would allow for proactive maintenance and prevent equipment damage or service disruptions. Another outcome from this artificial intelligence/machine learning (AI/ML) model was a partial short circuit compared to short circuit windings of an ideal transformer.

### Creating an Experimental Step-Down Transformer Model

A circuit simulation model was developed to capture primary and secondary data on a step-down ideal transformer. The electromagnetic component selected was an ideal transformer. The selection of such a component ensured that parasitics would not affect the transformer’s output secondary voltage, which would be degraded or produce harmonic behaviors. Figure 1 illustrates the transformer circuit model. A transient analysis simulation event was executed on the model to obtain primary and secondary voltage data from the step-down transformer.

##### Figure 1. Basic ideal transformer circuit model. Image used courtesy of Don Wilcher

A data table was built by varying the primary voltage of the transformer and measuring its secondary output voltage across the 10 KΩ resistive load. A sweeping voltage range of 120 Vac (169.71 Vp) to 10 Vac (3.03 Vp) was performed on the simulation model. Table 1 shows the collected data from the transformer circuit model. The model generated a two-cycle sinewave showing the primary and secondary voltages produced by the transformer circuit model.

##### Table 1. Primary and secondary voltages collected and recorded from the ideal transformer circuit model
 AC Voltage (V) Primary Voltage (Vpri) Secondary Voltage (Vsec) 120 169.71 33.94 110 155.56 31.11 100 141.42 28.28 90 127.28 25.46 80 113.14 22.63 70 99.99 19.99 60 85.85 17.17 50 71.71 14.34 40 57.57 11.51 30 43.42 8.68 20 29.28 5,86 10 15.14 3.03

Figure 2 illustrates the two-cycle sinewave produced by the AC electromagnetic component. The blue sinewave is the 169.71 peak voltage (Vp) applied to the ideal transformer’s primary winding, while the secondary winding’s 33.94 Vp is identified in red.

### Enter Machine Learning: The Power of Decision Trees

Machine learning offers powerful tools for analyzing complex data and identifying patterns. This project used a decision tree classifier, a popular supervised learning algorithm. A large language model (LLM) was used to create the Python code to establish the decision tree classifier. Specifications for partial short circuit, short circuit, and brownout behaviors were included during the prompt engineering session. Google Colaboratory was used to execute the AI-generated Python code to reduce library resources and dependencies on a standard laptop computer. A description of a decision tree is provided next.

Here's how decision trees work: Imagine a branching tree structure where each branch represents a decision based on a specific data feature. The primary input voltage and secondary output voltage are features in this laboratory exercise. The algorithm progressively "learns" by splitting the data based on these features, creating leaves at the end of each branch that represent a specific outcome—in this case, "normal operation" or "brownout." The decision tree classifier model observed partial short circuit or short circuit windings. Such conditions could be attributed to the step-down transformer's normal operation or brownout behaviors.

### Training the Model: Feeding the Data

A dataset containing voltage measurements and operational status (normal or brownout) for various input voltage levels trained the decision tree. The model analyzed these relationships, identifying patterns that differentiate normal operation from brownout conditions. The training process involves splitting the data into two sets. A training set is used to build the model; a testing set evaluates its performance. The model learns by iteratively adjusting its decision rules based on the training data.

### Evaluation: Testing the Model's Mettle

Once trained, the model is tested using the unseen data in the testing set. This assesses its ability to generalize the learned patterns and accurately predict brownout thresholds for new voltage inputs not encountered during training. Metrics like accuracy and precision are used to evaluate the model's performance. Accuracy measures the overall success rate in correctly predicting brownout conditions. Precision indicates how often a predicted brownout is a true positive (actual brownout). The model’s ability to identify actual brownout occurrences due to the partial short or short circuit transformer windings further aligns with the precision definition.

### AI/ML Potential for Predictive Power Grid Maintenance

The classifier’s decision tree results considered brownout conditions received by the transformer's primary windings. The classifier model provided a relational condition of this brown-out condition with primary voltage <= 107.5 V. Note that a brownout condition at 10% of the supply voltage of 120 Vac is 108. Based on the sample size and majority voting accomplished using a “DecisionTreeClassifier” ML model and supervised learning, the ideal transformer-partial short circuit winding was the most predictive outcome of the classifier. Based on this outcome, a brownout condition may be possible. This laboratory exercise, used with AI/ ML for predictive maintenance, illustrates potential analysis benefits in industrial and manufacturing applications. The results of the model are shown in Figure 3. A majority voting rule is used in a Decision Tree Classifier to predict the outcome of the fault detection element. In this laboratory exercise, the partial short circuit condition is the majority-ruled condition of the ideal transformer circuit model.

##### Figure 3. Results of the Decision Tree Classifier model. Image used courtesy of Don Wilcher

Finally, this laboratory exercise demonstrates the potential of AI/ML for predictive maintenance in power grids, contributing to a more resilient and efficient power infrastructure by predicting brownout thresholds in step-down transformers. As research continues, further advancements in model complexity, real-time monitoring capabilities, and integration with existing grid management systems can unlock the full potential of AI for a reliable and sustainable power supply.