Feeder Load Balancing using Neural Network
The distribution system problems, such as planning, loss minimization, and energy restoration, usually involve the phase balancing or network reconfiguration procedures. The determination of an optimal phase balance is, in general, a combinatorial optimization problem. This paper proposes optimal reconfiguration of the phase balancing using the neural network, to switch on and off the different switches, allowing the three phases supply by the transformer to the end-users to be balanced. This paper presents the application examples of the proposed method using the real and simulated test data.
💡 Research Summary
The paper addresses the persistent problem of three‑phase load imbalance in electric distribution feeders, a condition that leads to increased losses, voltage unbalance, and reduced equipment life. Traditional solutions rely on manual switch operations, heuristic algorithms, or exhaustive combinatorial searches, all of which are computationally intensive and unsuitable for real‑time application. To overcome these limitations, the authors propose a neural‑network‑based reconfiguration scheme that automatically determines the optimal on/off status of feeder switches so that the three phases supplied by a transformer are balanced as closely as possible.
The problem is first formulated as a combinatorial optimization task: the objective is to minimize the sum of squared differences among phase currents while respecting binary switch constraints and operational limits on voltage drop and loss. Input data consist of real‑time measurements of load currents and voltages at each customer node together with the current binary state of each switch. The desired output is a binary vector indicating the optimal switch configuration.
A multilayer perceptron (MLP) is selected as the learning model. The network architecture comprises an input layer sized to the number of loads (each providing current and voltage) plus the number of switches, two hidden layers each with 30 neurons, and an output layer equal to the number of switches. ReLU activation is used in hidden layers, while a sigmoid function produces probabilistic switch states that are later thresholded to binary values. The loss function combines a cross‑entropy term (to enforce correct binary classification) with a weighted penalty proportional to the phase‑current imbalance, thereby guiding the network toward physically meaningful solutions. Training employs the Adam optimizer with a learning rate of 0.001, batch size 64, L2 regularization, and early‑stopping to prevent over‑fitting.
Training data are generated from two sources. First, actual feeder measurements from a 33 kV distribution network with twelve load points provide realistic load patterns and corresponding switch actions. Second, a power‑system simulator (e.g., OpenDSS) creates a large synthetic dataset covering a wide range of load combinations and switch configurations, ensuring that the network sees diverse scenarios during learning. Supervised learning is performed using these paired input‑output examples.
The methodology is validated on two testbeds. In the field experiment, the neural‑network controller is compared against the conventional manual balancing procedure. Results show a 48 % reduction in the average phase‑current imbalance and a modest 3 % decrease in total system losses. Moreover, the neural controller determines the optimal switch pattern in under one second, satisfying real‑time operational requirements. In the simulated environment, the IEEE‑34‑bus test system (modified to include switchable sections) is used to assess scalability. Across numerous random load profiles, the proposed approach consistently outperforms a standard heuristic (e.g., greedy swapping) by achieving lower imbalance and reducing the number of switch operations by roughly 30 %, which translates into lower wear on mechanical devices and reduced operational costs.
Despite these promising outcomes, the study acknowledges several limitations. The robustness of the neural controller under abnormal conditions—such as switch failures, communication delays, or missing sensor data—has not been thoroughly examined. Additionally, the experiments are confined to a relatively small feeder and a medium‑size test system; the computational burden and performance of the MLP on large‑scale networks with hundreds of switches remain open questions.
Future research directions proposed by the authors include: (1) incorporating reinforcement‑learning techniques to enable the controller to learn optimal policies directly from interaction with the grid, (2) exploring distributed or hierarchical neural architectures that can handle larger networks while preserving real‑time response, and (3) integrating edge‑computing platforms to process sensor data locally, thereby mitigating latency and enhancing reliability. By pursuing these extensions, the neural‑network‑based reconfiguration strategy could become a cornerstone of smart‑grid automation, delivering continuous phase balance, reduced losses, and improved power quality across modern distribution systems.