Indirect Short-Term Net Load Forecasting Using Ensemble Learning and Anomaly

Detection for Solar Integrated Microgrids

Papers Covered
1. Direct Against Indirect Short-Term Net Load Forecasting Using Machine Learning
Principles for Renewable Microgrids
Publisher
2023 IEEE International Smart Cities Conference (ISC2)
Year: 2023 | Conference Paper | Publisher: IEEE

This paper investigates and compares the performance of direct and indirect strategies for
short-term net load forecasting (STNLF) in renewable microgrids using Bayesian Neural
Network (BNN) models.
Key Findings:
 Direct STNLF outperforms indirect STNLF: The direct STNLF strategy consistently
demonstrated lower RMSE and nRMSE values compared to the indirect strategy for
buildings and the University of Cyprus (UCY) microgrid.
 This superior performance was particularly pronounced in cases with high PV
penetration.
 Optimal BNN model structure: The research identified optimal BNN model structures
for both direct and indirect STNLF:
 Direct: 8 input nodes (historical net load, month, day, dew point temperature, real feel
temperature, ambient temperature, time of day, and global horizontal irradiance), 7
hidden nodes, and 1 output node (net load).
 Indirect: Separate BNN models for load and PV generation forecasting were
developed and their outputs combined.
 Load forecasting: 9 input nodes (historical load, global horizontal irradiance, ambient
temperature, real feel temperature, month, relative humidity, day, dew point
temperature, and time of day) and 7 hidden nodes.
 PV generation forecasting: 6 input nodes (historical PV generation, global horizontal
irradiance, ambient temperature, real feel temperature, relative humidity, and time of
day) and 5 hidden nodes.
  Importance of data quality and feature selection: The study emphasizes the importance
of data quality assessment and careful input feature selection using correlation analysis
and mutual information to ensure accurate forecasting.
 "To ensure high data quality, the raw datasets were subjected to a series of
filtering/deletion and inference stages, to identify duplicates, incorrect measurements,
and missing values."
Key Facts & Figures:
 The UCY microgrid has a maximum measured net load power of 1645.77 kW and a
PV capacity of 434.80 kWp, representing a penetration share of 26.42%.
 Direct STNLF achieved nRMSE values of 4.81%, 5.35%, and 3.98% for the social
facilities building (SFB), faculty science building (FSB), and the UCY microgrid,
respectively.
 Indirect STNLF resulted in nRMSE values of 5.80%, 5.38%, and 4.17% for the SFB,
FSB, and UCY microgrid, respectively.
Implications:
 The findings advocate for the adoption of direct STNLF for renewable microgrids,
particularly those with high PV penetration, due to its superior accuracy.
 The research provides a framework for developing and optimizing BNN models for
both direct and indirect STNLF, which can be adapted to other microgrids.
Overall: This study provides valuable insights into STNLF strategies for renewable
microgrids, demonstrating the advantages of a direct approach and offering practical
guidance for developing accurate forecasting models.

Research gap
The research gap this paper addresses is the lack of comparative studies on direct and
indirect short-term net load forecasting (STNLF) strategies specifically for renewable-based
microgrids.
While previous studies have compared these strategies in residential, distribution, and power
systems, this paper focuses on applying them to microgrids with high renewable penetration.
This is important because the increasing share of renewables in microgrids necessitates
accurate NLF for efficient operation and management.
The authors aim to fill this gap by comparing the performance of direct and indirect STNLF
strategies based on Bayesian neural network (BNN) principles using historical data from the
University of Cyprus (UCY) microgrid and buildings.
Future Scope
Implementation and evaluation of indirect and direct STNLF strategies in larger capacity
microgrids with higher renewable shares.
Incorporation of the direct STNLF approach into web-based forecasting tools.
2. LSTM-Based Net Load Forecasting for Wind and Solar Power-Equipped
Microgrids
Publisher
2024 56th North American Power Symposium (NAPS)
Year: 2024 | Conference Paper | Publisher: IEEE
General Review
This research paper proposes and evaluates a deep learning model for net load forecasting in
microgrids incorporating solar and wind power.
The model uses Long Short-Term Memory (LSTM) networks to predict net load—the
difference between electricity demand and renewable energy generation—leveraging a
typical meteorological year (TMY) dataset.
The authors compare a "direct" forecasting approach against an "indirect" approach, where
renewable generation and demand are predicted separately before calculating net load.
The study finds the indirect method superior, achieving a normalized root mean squared
error (nRMSE) of 9.4%, demonstrating the effectiveness of the LSTM model for enhancing
energy management in renewable-based microgrids.
The paper details the model's architecture, data sources, training methodology, and results,
concluding that the indirect approach offers improved prediction accuracy.
Mathematical modeling
Detailed mathematical modeling of solar and wind power generation was introduced
Long Short-Term Memory (LSTM) Model
The LSTM model demonstrates remarkable effectiveness in predicting net load for
renewable-based microgrids.
The direct approach involves training the model directly with derived net load data.
The indirect approach trains the model separately to predict solar power, wind power, and
load demand, combining these predictions to calculate the net load.
Model Training
The LSTM-based model was trained using data from three primary sources:
1. Hourly ambient temperature and wind speed data for various locations in Texas for the
year 2019
2. Global horizontal irradiance (GHI), direct horizontal irradiance (DHI), and direct normal
irradiance (DNI) data for a typical year across multiple locations in the United States,
including Texas.
3. Average load data for residential units for a typical year in the same locations as the solar
irradiance data.
The data from sources1 and2, which includes wind speed, temperature, and irradiance, was
used to derive the solar and wind power output of the microgrid using equations described in
the methodology section of the source document. This derived data was combined with the
load demand data from source3 to create the net load labels for training the LSTM model.
The input dataset for both the direct and indirect approaches consisted of 8760 data points
(24 hours x 365 days) with five columns:
Day number
Hour of the day
Temperature (K)
Wind speed (m/s)
Collector irradiance (W/m²)
The output dataset for both approaches contained 8760 data points with one column,
representing net load in the direct approach, or load demand, wind power, and solar power
separately in the indirect approach.
The input data was normalized and partitioned into training (80%), validation (10%), and
testing (10%) datasets to improve the training process and allow for model evaluation.
Model Architecture
The proposed deep learning (DL) model for net load forecasting (NLF) is based on Long
Short-Term Memory (LSTM) networks, a type of recurrent neural network specifically
designed to handle time series data and learn long-term dependencies.
The specific architecture of the LSTM-based model implemented for this study involves:
Three LSTM layers: These layers take the input data (day number, hour of the day,
temperature, wind speed, and collector irradiance) and learn the temporal relationships
between the input features and the net load labels.
Regularization: Each LSTM layer uses L2 regularization with a lambda value of 0.001 to
prevent overfitting.
Dropout: Dropout rates of 0.4 are applied to further prevent overfitting.
Batch normalization: This technique is employed to stabilize the learning process and
accelerate training.
Dense layer with linear activation: This layer follows the LSTM layers to learn further
correlations between input data and labels, preserving potential negative values in the net
load prediction (which can occur when renewable generation exceeds load demand).
Output layer: This final layer condenses all the learned information to produce a single
output: the predicted net load.
The researchers tested this LSTM-based model using both direct and indirect approaches to
NLF. The direct approach uses the model to predict the net load directly, while the indirect
approach uses the model to predict solar power, wind power, and load demand separately,
and then combines those predictions to calculate the net load. Both approaches use the same
input dataset of meteorological and time-related data, as described in our previous
conversation.
Metrics to evaluate the performance
Mean Absolute Error (MAE): This metric measures the average absolute difference
between the predicted and actual net load values.
Mean Squared Error (MSE): This metric calculates the average squared difference
between the predicted and actual values, giving more weight to larger errors.
Root Mean Squared Error (RMSE): The square root of the MSE, providing an error
metric in the same units as the net load.
Normalized Root Mean Squared Error (nRMSE) This metric expresses the RMSE as a
percentage of the average actual net load, allowing for comparison across different datasets.
Future Scope
Improving the model's ability to handle extreme weather events and anomaly
3. A Comprehensive Review of Load Demand Forecasting and Optimal
Sizing In Stand Alone Hybrid Renewable Energy Systems
Publisher
2023 Annual International Conference on Emerging Research Areas: International
Conference on Intelligent Systems (AICERA/ICIS)
Year: 2023 | Conference Paper | Publisher: IEEE
 Importance of Load Demand Forecasting: Accurate load forecasting is critical for
planning and managing energy systems, ensuring supply meets demand and
optimizing resource utilization.
 Optimal Sizing for Hybrid RES: Choosing the right size and combination of
renewable energy sources (e.g., solar PV, wind turbines) and storage technologies is
essential for cost-effective and reliable off-grid power systems.
 Factors Influencing Load Demand: Load demand is affected by numerous variables
including weather, seasonal variations, time of day, and consumer behavior.
Understanding these factors is essential for accurate forecasting.
 Forecasting Methods: Various techniques are employed, ranging from statistical
methods (e.g., time-series analysis) to artificial intelligence (AI) approaches (e.g.,
deep learning, machine learning).
 Challenges and Limitations: Data accuracy, model complexity, and uncertainties in
renewable energy availability pose challenges to forecasting and optimal sizing.
Important Ideas/Facts:
 Deep Learning for Load Forecasting: Deep learning models, like ANNs, RNNs,
LSTMs, and GRUs, are increasingly utilized due to their ability to handle complex
temporal relationships and achieve high forecasting accuracy However, reliance on
future weather data can impact accuracy due to inherent uncertainties
 Machine Learning Applications: Machine learning techniques, including SVR,
DNN, XGBoost, and Q-learning, are also being explored for forecasting heat load,
PHEV charging patterns, and optimizing energy resource utilization
  Hybrid Approaches: combine AI and statistical methods to leverage the strengths of
both. For example, a study used a hybrid DL model combining LSTM and CNN for
short-term load forecasting and achieved good results.
 Transfer Learning Potential: Transfer learning shows promise in reducing
computational costs and improving forecasting accuracy by applying pre-trained
models to new contexts.
 Optimal Sizing Considerations: Factors influencing optimal sizing include
renewable energy resource availability, load energy demand, storage capacity, and
overall system cost.
 Benefits of Hybrid RES: Hybrid RES can increase renewable energy usage, reduce
reliance on fossil fuels, and contribute to a cleaner environment.

AI-based load forecasting methods, such as Artificial Neural Networks (ANNs), fuzzy logic,
and Support Vector Machines (SVMs), are generally considered to have high accuracy
compared to traditional statistical methods like time-series analysis and regression analysis.

Deep Learning (DL) models, a subset of AI methods, are particularly effective due to their
ability to learn temporal relationships and consider multiple input variables.

Bayesian DL models, showed its effectiveness in handling the high uncertainty and volatility
of residential load forecasting.

However, the accuracy of AI-based methods can be affected by factors such

as:
 Data quality and availability: Building-level load data might be limited, affecting the
training effectiveness of DL models.
 Model complexity: Complex problems often require repeated evaluation of the fitness
function, which can be computationally expensive.
 Uncertainty and generalization: ELM, another AI-based method, faces challenges
related to uncertainty and generalization degradation, which can impact accuracy.

Statistical methods, such as time-series analysis and regression analysis, rely on historical
data to identify patterns and trends to make predictions1. These methods are generally
simpler to implement and computationally less expensive than AI-based methods.
However, they may struggle to capture complex non-linear relationships in load data and
are sensitive to noise and seasonal trends.

Deep Learning (DL) and Machine Learning (ML) in Load Forecasting

Deep Learning (DL) and Machine Learning (ML) are both subfields of Artificial Intelligence
(AI) used for load forecasting, but they differ in their approach and capabilities.

Machine Learning:

 ML algorithms learn from data to make predictions without explicit programming.

 They can handle complex problems but may require manual feature engineering.
 Examples of ML algorithms used for load forecasting include Support Vector
Machines (SVMs), Extreme Learning Machines (ELMs), and Fuzzy C Means.

Deep Learning:

 DL is a subset of ML that uses artificial neural networks with multiple layers to learn
complex representations of data.
 DL models can automatically learn features from data, reducing the need for manual
feature engineering.
 They are particularly effective for load forecasting due to their ability to learn
temporal relationships and consider multiple input variables.
 Examples of DL algorithms used for load forecasting include Artificial Neural
Networks (ANNs), Recurrent Neural Networks (RNNs), Convolutional Neural
Networks (CNNs), Long Short-Term Memory (LSTM) networks, and GRU.

Key Differences:

 Feature Engineering: ML often requires manual feature engineering, while DL can

automatically learn features from raw data.
 Data Requirements: DL typically requires larger datasets than ML for effective
training.
 Complexity: DL models are generally more complex and computationally intensive
than ML models.
 Interpretability: ML models are often more interpretable than DL models, which can
be considered "black boxes."

Choosing Between DL and ML:

The choice between DL and ML depends on factors like data availability, computational
resources, and the complexity of the forecasting problem.
In load forecasting, DL has shown promising results due to its ability to handle the complex
non-linear relationships and temporal dependencies in load data. However, ML methods can
still be effective for specific applications, especially when data is limited or computational
resources are constrained.

Spatio-temporal correlation
Spatio-temporal correlation in power systems refers to the relationship between spatial
(geographical) and temporal (time-related) variations in power generation, consumption, and
system performance. Understanding these correlations is crucial for optimizing the operation
and planning of power systems, especially with the increasing integration of renewable
energy sources like wind and solar, which can be highly variable both in space and time.
Here are a few key points about spatio-temporal correlation in power systems:
Load Forecasting: By analyzing historical data, utilities can identify patterns in electricity
demand across different regions and times. This helps in predicting future load requirements
more accurately.
Renewable Energy Generation: The output from renewable sources can vary significantly
based on location and time of day. For instance, solar power generation is highest during
sunny days, while wind energy can fluctuate based on weather conditions. Understanding
these correlations helps in better integrating renewables into the grid.
Demand Response Programs: Utilities can use spatio-temporal correlations to design demand
response programs that encourage consumers to adjust their usage based on real-time
conditions, helping to balance supply and demand.
In summary, spatio-temporal correlation is a vital concept in modern power systems,
enabling better forecasting, integration of renewable energy, and overall system reliability
Future Scope
Optimizing Energy Storage Systems: Efficient energy storage is crucial for balancing the
intermittent nature of renewable energy sources. Further research is needed to develop
advanced energy storage technologies with higher capacity, faster response times, and longer
lifespans
Integrating Demand Response: Demand response programs encourage consumers to adjust
their energy consumption patterns based on price signals or other incentives
4. Direct Short-Term Net Load Forecasting Based on Machine Learning
Principles for Solar-Integrated Microgrids
Publisher
Year: 2023 | Volume: 11 | Journal Article | Publisher: IEEE

This paper proposes a novel methodology for directly forecasting short-term net load in
solar-integrated microgrids using Bayesian Neural Networks (BNNs).
The study emphasizes the importance of accurate net load forecasting for managing the
increasing integration of variable renewable energy sources (RES) like solar PV.
Most Important Ideas & Facts:
 Direct NLF: The authors advocate for a direct net load forecasting (NLF) approach,
where a single model predicts the net load (aggregated load minus RES production)
instead of forecasting load and generation separately. This offers computational
advantages and doesn't require detailed knowledge of system characteristics or behind-
the-meter PV installations.
 BNN Model: BNNs are chosen due to their ability to capture uncertainty, handle non-
linear relationships, and prevent overfitting. The study details the construction of the
BNN model, including input feature selection and optimization.
 Key features: Historical net load, time-related parameters (month, day, time), and
weather data (GHI, dew point temperature, real feel, ambient temperature) are
identified as crucial inputs.
 Optimal architecture: The optimal model uses 8 input features, 7 hidden nodes, and
1 output node.
 Model Training: The study explores the impact of training data size and split
approach (sequential vs. random) on model performance. Randomly splitting the data
with a 70:30 train-test ratio achieves the best results.
 Performance Evaluation: The model is rigorously evaluated using multiple metrics
(nRMSE, RMSE, Skill Score) on a dataset from a university campus microgrid with
varying PV penetration levels.
 Accurate Results: The optimal BNN model achieves a normalized root mean square
error (nRMSE) of 3.98% for the entire microgrid, outperforming a baseline naive
persistence model (NPM).
  Robustness: The model's robustness is validated using k-fold cross-validation and an
unseen three-month dataset, consistently yielding low nRMSE values.
 Benefits for Decision Making: Accurate STNLF can significantly improve decision-
making for microgrid operators, enabling optimized energy management, grid
stability, and reliable integration of solar PV.
Future Scope

Incorporation of Extreme Events:

Incorporating additional input features that capture such events or using more advanced
modeling techniques that are less sensitive to outliers.