Deep Fault Analysis and Subset Selection in Solar Power Grids

• Renewable energy can be wasted through mismanagement, storage problems, and non-optimal generation choices, while faults and inaccurate load forecasts contribute to outages and congestion.Source: §1 · p. 1
• The paper proposes LSTMs to analyze grid information and determine grid health during faults.Source: §1 · p. 1
• It also uses Siemens PSS/E simulations of generator supply, weather, and load demand to select optimal solar-generator subsets.Source: §1 · p. 1

2.1 Data

• Grid data were generated with Siemens PSS/E and PowerWorld Simulator, with PSS/E supporting power-flow and dynamic or transient analysis.Source: §2.1 · p. 2

2.2 Determining the fault type

• Experiments used a 23-bus grid with four simulated fault types, running each fault at every bus 100 times to generate variations.Source: §2.2 · p. 2
• The fault-type model processes 100 time-unfolded LSTM voltage inputs and uses temporal information to classify LL versus LG faults.Source: §2.2 · p. 2
• 95% LSTM accuracy for fault classification improved approximately 6% over the baseline SVM model.Source: §2.2 · p. 2

2.3 Deducing the faulted bus line number

• After fault-type classification, separate LSTM models predict the triggered bus number or output 0 for non-faulty network data.Source: §2.3 · p. 2
• 97% accuracy was reported for predicting the faulted bus number, including outputting 0 for non-faulty data.Source: §2.3 · p. 2, Figure 1 · p. 2

• The proposed technique aims to reduce renewable power-forecasting uncertainty while accounting for congestion and load demand in economic dispatch.Source: §3 · p. 3
• Solar generators face imbalance penalties because their contracted power may differ from actual output and traditional generators are generally more reliable.Source: §3 · p. 3

• Solar-power data were collected for 16 days, from March 1 through March 14, 2017, using rooftop panels at IIT Kharagpur’s Electrical Engineering department.Source: §3.1 · p. 3

3.2 Congestion prediction

• 97–98% test accuracy was achieved when actual generated solar power was known, while unknown-power prediction reached around 91%.Both neural-network models improved accuracy by around 5% over the baseline SVM model.Source: §3.2 · p. 3

3.2.1 Data Generation for actual subset selection

• The data-generation process switched off generator subsets after an initial all-solar load flow to observe one-step grid-state changes and congestion.The objective was to meet demand at the least cost using solar and non-conventional generators.Source: §3.2.1 · p. 3

3.3 Choosing the optimal subset

• The optimal subset is selected by minimizing total penalty, combining power mis-commitment and congestion penalties.The objective is to choose solar generators whose operation incurs the least overall penalty.Source: §3.3 · p. 3

3.3.1 Power mis-commitment penalty

• Weather uncertainty creates differences between committed and delivered power, so the mis-commitment objective favors subsets with lower prediction variance.Larger committed-versus-delivered power differences incur greater penalties for grid operators.Source: §3.3.1 · p. 3

3.3.2 Penalty incurred due to congestion in the network

• A congestion penalty is added because a subset with lower mis-prediction cost may still cause undesirable network congestion.The congestion loss is L2 = 50 when a subset causes congestion and L2 = 0 otherwise.Source: §3.3.2 · pp. 3–4

3.4 The Model

• Inputs combine bus voltages before subset selection with a binary vector indicating which solar generators belong to the subset.For each possible subset, the total loss is computed as L1 + L2 and used as the target output.Source: §3.4 · p. 4
• The model is trained to output and minimize the total loss incurred by selecting the input subset.The loss function is defined from the model output and the combined penalty target.Source: §3.4 · p. 4
• The network uses one hidden layer with 200 hidden neurons and is trained with the Adam optimizer.Adam is used to minimize the loss L.Source: §3.4 · p. 4

3.5 Results

• Training minimum L2 loss was 35 and L1 loss was 6, while test L2 loss was 59 and L1 loss was 8 after around 2500 steps.Figure 2 shows the training and test L2-loss plots.Source: §3.5 · p. 4, Figure 2 · p. 4
• The authors report that the low test losses indicate selection of an economical subset with the least penalty and lowest possible congestion probability.Source: §3.5 · p. 4

• The paper contrasts its LSTM representation of time-series voltage and angle data with ANN and SVM methods for locating faults in radial distribution systems.Source: §4 · p. 4
• Unlike modified PSO methods for handling equality and inequality constraints in economic dispatch, this approach uses machine learning and deep learning to choose generator subsets.Source: §4 · p. 4

• The proposed system predicts congestion and selects economical generator subsets, including configurations with or without solar generators, using deep learning.Source: Conclusion · p. 4
• Knowledge of grid vulnerability is presented as a basis for handling load shedding and power surges efficiently.Source: Conclusion · p. 4
• The work is positioned as a step toward power security where developing countries require clean and reliable electricity for rapid development.Source: Conclusion · p. 4