Distribution system monitoring for smart

power grids with distributed

generation using artificial neural network
Jan-Hendrik Menke, Nils Bornhorst, Martin Braun Department of Energy Management and Power System Operation,
University of Kassel, Wilhelmshöher Allee 71-73, 34121 Kassel, Germany

ABSTRACT

This paper proposes an innovative ANN-based monitoring

scheme for distribution grids, overcoming significant limitations
of existing methods. Current approaches often rely on static
grid topologies, only estimate voltage magnitudes, and struggle
with grids featuring high levels of distributed generation, such
as photovoltaic and wind energy sources. The proposed method
accurately estimates both voltage magnitudes and line loadings
while requiring far fewer measurements than traditional
weighted least squares state estimation (WLS SE). It
accommodates dynamic switching states and employs a
scenario generator to produce diverse, robust training data.
Extensive simulations on real and benchmark grids
demonstrate its superior accuracy, scalability, and resilience
under various operating conditions, including scenarios with
gross measurement errors. This work sets a new standard for
efficient and reliable grid monitoring in modern power systems.

1. Introduction

The increasing integration of distributed generation (DG),

such as photovoltaics (PV) and wind energy converters
(WEC), in medium and low-voltage grids has made power
flows increasingly unpredictable. Unlike conventional power
plants, DG sources have volatile power injections, causing
significant variations in both the magnitude and direction of
power flow. To delay costly grid reinforcements, grids are
 often operated closer to their operational limits, creating a
need for precise monitoring to prevent violations. Traditional
grid monitoring methods like weighted least squares state
estimation (WLS SE) rely on a large number of
measurements, but this approach is often uneconomical for
distribution grids due to the expensive infrastructure
required.

A novel artificial neural network (ANN)-based monitoring

scheme is proposed to overcome the limitations of existing
methods. The approach addresses common issues such as
low estimation accuracy, high computational efforts, and the
inability to account for dynamic switching states. By
generating controlled training data using a scenario
generator and estimating both voltage and line current
magnitudes, the scheme provides accurate monitoring of
grids with high DG penetration. This method requires fewer
measurements and reduces reliance on complex
infrastructure. The effectiveness of the scheme is
demonstrated through simulations on benchmark grids and
real-world distribution grids with high DG penetration,
proving its practicality for modern grid monitoring.
2. Artificial neural network
An Artificial Neural Network (ANN) is a computational system
composed of interconnected elements known as neurons, which
exchange information through weighted connections. A perceptron, a
widely used neuron model, processes multiple inputs ((x_1,
x_2, ..., xn)), applies weights ((wi)), sums them up, and passes the
result through an activation function like the Rectified Linear Unit
(ReLU): (y = f(x) = max(0, sum wx)). The perceptron's output forms the
simplest neural network. Typically, ANNs consist of multiple
perceptrons organized into layers, with outputs from one layer serving
as inputs to the next, forming a multi-layer perceptron (MLP). ANNs are
trained using data through iterative epochs. During each epoch, the
training dataset adjusts the weights using an optimizer to match
desired outputs. Validation datasets help prevent overfitting by
stopping training at an optimal point. The presented scheme involves a
preparation phase (training the ANN using generated data scenarios)
and an operation phase where live measurements are used to predict
 desired values efficiently. This approach ensures accurate predictions
for real-time applications.

Figure 1 . Flow chart of the scheme.

3. Grid monitoring with ANN

Monitoring electrical power grids using Artificial Neural Networks
(ANNs) addresses challenges in distribution grids with sparse
measurements and high Distributed Generation (DG) penetration.
Unlike traditional Weighted Least Squares State Estimation (WLS SE),
ANNs function as data-driven models, learning to map limited
measurement inputs to desired grid variables, such as bus voltages
and line currents.A multi-layer perceptron architecture is employed,
optimized with three hidden layers, ReLU activation functions, and the
Adam optimizer. Input data includes measurements and switching
states, while outputs represent bus voltage magnitudes or line current
magnitudes. Training relies on scenarios generated to simulate diverse
grid conditions, incorporating Gaussian noise to model device tolerance
and improve generalization. This scenario-based training ensures the
ANN captures variability in load, photovoltaic (PV), and wind energy
converter (WEC) power. The proposed method demonstrates
robustness under real operating conditions, accounting for topology
changes, data loss, and time constraints (e.g., sub-second runtimes).
Scenario generation effectively spans operational ranges, ensuring
high accuracy in both training and validation. Future extensions could
enable applications in unbalanced grids by scaling input and output
variables per phase or incorporating voltage angle estimation for full
state estimation. This adaptable, efficient approach offers significant
improvements for monitoring modern, dynamic distribution grids.
Figure 2 Data flow into and out of the ANN.

4. Electrical test network

The ANN-based grid monitoring accuracy was evaluated using two test
grids:
1. CIGRE MV Benchmark Grid: This 20 kV grid has 15 buses, 8 PV
generators, and 1 WEC. Some DG units were omitted, and the
remaining DG power was doubled to create more challenging
conditions. Four switching configurations were analysed.
2. Real Distribution Grid in Germany: A 20 kV radial grid with 135
buses, 134 lines, 82 loads, 75 PV generators, and 2 wind farms. It
features 84 measurements (P, Q, V) across 19 buses. The grid’s DG
power is nearly three times the load. Five switching configurations
were tested.
Training data, with IEC 61869-compliant noise, was generated using a
scenario generator and PFC. Accuracy classes (ACC) 0.5 and 1 were
used for voltage and current measurements, respectively. Simulations
assessed ANN monitoring viability.
5. Simulations
The implementation of the artificial neural network (ANN) for
monitoring and state estimation in power grids is carried out using
Python, with the PyTorch and pandapower packages. State estimation
(SE) is based on the weighted least squares (WLSSE) method, with a
maximum iteration count set to 10 and a stopping criterion based on
state update thresholds. Substitute (or pseudo) measurements are
generated for unmeasured bus power injections by distributing
generated power based on the installed power of the loads at each bus.
These pseudo-measurements are adjusted with a standard deviation of
30% to account for their imprecise nature.
The method is evaluated under various operational conditions, such as
normal operations and cases involving measurement errors or topology
issues. The monitoring performance is assessed using two evaluation
criteria (C1 and C2), where successful estimations are those with
 estimation errors within predefined limits. Different test cases with
varying measurement configurations are used to assess the robustness
of the method under errors like measurement outages or gross errors.
In terms of computational resources, the implementation was run on an
Intel i7-4702MQ CPU with 16 GB RAM and SSD storage, using Python
3.6 on Ubuntu Linux, without GPU acceleration.
6. Conclusions
An ANN-based scheme is developed for reliable distribution system
monitoring with minimal measurements, addressing limitations of
existing methods in grids with high DG penetration. A scenario
generator ensures diverse training data, enabling accurate
performance across varied grid states. Validated on benchmark and
real grids, the approach consistently outperforms WLS state estimation
(WLS SE), even with identical input data. The method extends to
imbalanced grids by processing phases separately or training with
multi-phase data. Future directions include improving accuracy under
bad data conditions, optimizing training data for complex grids,
automating evaluation, and reducing training time using parallel ANN
training.
7. Acknowledgement
I would like to express my heartfelt gratitude to my supervisor, Pankaj
Mukhija Sir, for their expert guidance and unwavering support
throughout this research on ANN-based monitoring of distribution grids.
I extend my thanks to the Electrical Engineering Department at NIT
Delhi for providing the necessary resources and infrastructure. I am
deeply appreciative of my colleagues for their valuable insights and
assistance during the project. Finally, I thank my family and friends for
their constant encouragement and motivation, which have been
instrumental in the completion of this work.
8. References
1. Abur, A., & Expósito, A. *Power System State Estimation: Theory and
Implementation.* CRC Press, 2004.
2. Mosbah, H., & El-Hawary, M. "Power system tracking state estimation
based on stochastic fractal search technique under sudden load changing
conditions." *IEEE Canadian Conference on Electrical and Computer
Engineering (CCECE) ([Link]
3. Filho, M.B.D.C., & de Souza, J.C.S. "Forecasting-aided state estimation;
Part I: Panorama." *IEEE Transactions on Power Systems
([Link]
4. Menke, J.-H., Hegemann, J., Gehler, S., & Braun, M. "Heuristic
monitoring method for sparsely measured distribution grids."
*International Journal of Electrical Power & Energy Systems, (
[Link]
5. Manitsas, E., Singh, R., Pal, B.C., & Strbac, G. "Distribution system state
estimation using an artificial neural network approach for pseudo
measurement modeling." *IEEE Transactions on Power Systems,
([Link]