Paper 1

The paper titled "Hybrid CNN-LSTM Model for Short-Term Individual Household Load
Forecasting" proposes a deep learning model combining Convolutional Neural Networks
(CNN) and Long Short-Term Memory (LSTM) networks to improve the accuracy of short-
term electric load forecasting for individual households.

Methods
Data Collection and Feature Preparation: Data was sourced from the Smart Grid Smart
City (SGSC) project, focusing on half-hourly household power consumption. Features
included historical consumption data, time of day, day of the week, and holiday
indicators.

CNN-LSTM Model:

CNN Layers: These layers extract features from the input data and filter out noise,
helping the model identify relevant patterns in time-series data.
LSTM Layers: These capture sequential dependencies, allowing the model to recognize
both short-term and long-term patterns.
Architecture: The model architecture includes three convolutional layers for
feature extraction, followed by LSTM layers for sequence learning. The model’s
final layers produce a forecast for household energy demand at different time
intervals.
Evaluation Metrics: The model was evaluated using Mean Absolute Percentage Error
(MAPE) and compared against other models such as simple LSTM and traditional
machine learning models.

Results
Single-Step Forecasting: The proposed model achieved a MAPE of 40.38% for a 2-time-
step lookback, outperforming the standard LSTM model, which had a MAPE of 44.06%.

Multi-Step Forecasting: For forecasting up to three hours ahead, the CNN-LSTM model
showed consistent improvements. For instance, with a 2-time-step lookback, it
achieved 4.01% lower MAPE for one-time step forecasting and up to 5.98% improvement
over the LSTM model for six-time steps ahead.

Paper 2

The paper "Optimised Deep Learning for Time-Critical Load Forecasting Using LSTM
and Modified Particle Swarm Optimisation" proposes a novel hybrid model that
combines Long Short-Term Memory (LSTM) with Modified Particle Swarm Optimization
(mPSO) to improve accuracy, stability, and convergence rate in load forecasting.
Here’s an overview of the methods and results:

Methods
Hybrid Model:

The model integrates LSTM for handling time-series data, which is optimized by the
mPSO algorithm to fine-tune parameters like learning rate, dropout rate, and batch
size.
The LSTM is enhanced by a Hybrid Feature Selection (HFS) module, which combines
XGBoost, Decision Trees (DT), and Recursive Feature Elimination (RFE) to select the
most relevant features, reducing data dimensionality and enhancing forecasting
accuracy.
Modified Particle Swarm Optimization (mPSO):
mPSO dynamically adjusts the cognitive and social coefficients, as well as inertia
weight, for better convergence and global optima avoidance. This tuning capability
allows efficient exploration and optimization in the parameter space, ensuring
robust LSTM training.
Comparison with Benchmark Models:

The model’s performance was compared to existing frameworks like Elman Neural
Network (ENN), conventional LSTM, Genetic Algorithm-optimized LSTM (GA-LSTM), and
PSO-LSTM models.
Evaluation Metrics:

Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute
Percentage Error (MAPE) were used to evaluate the model’s accuracy.
Results
Performance Improvements:

The proposed HFS-LSTM-mPSO model achieved a significant reduction in MSE, RMSE, and
MAPE for both NSW and VIC datasets when compared to benchmark models.
For NSW, the model reduced MSE by 91.91%, RMSE by 94.89%, and MAPE by 74.29%. For
VIC, it achieved MSE reduction by 91.33%, RMSE by 95.73%, and MAPE by 72.06%.

Paper 3
The paper “Load Forecasting with Machine Learning and Deep Learning Methods”
evaluates several AI methods—random forest (RF), support vector regression (SVR),
extreme gradient boosting (XGBoost), multilayer perceptron (MLP), long short-term
memory (LSTM), and temporal convolutional network (Conv-1D)—for forecasting
electrical loads in buildings. Here is a summary of the methods and results:

Methods
Data Preprocessing: The data was cleaned, scaled, and prepared for pattern
recognition, with missing values addressed through interpolation. Data included
occupancy, outdoor temperature, and temporal features (hour, day, and season).

Modeling:

Machine Learning (ML) models: RF, SVR, and XGBoost were employed. RF aggregates
multiple decision trees for generalization, SVR fits data within a boundary range,
and XGBoost optimizes tree-based learning through gradient boosting.
Deep Learning (DL) models: MLP, LSTM, and Conv-1D were used. MLP is a feedforward
neural network with fully connected layers; LSTM is a recurrent neural network
tailored for time-series data, capturing long-term dependencies; Conv-1D employs
convolutional layers to extract temporal features.
Evaluation Metrics: Model performance was assessed using normalized mean bias error
(nMBE) and normalized root mean squared error (nRMSE). Cross-validation (10-fold)
was applied to ensure robustness.

Results
Bias (nMBE): In the validation set, LSTM had the lowest bias, with an nMBE of -
0.02%, indicating minimal overestimation, while SVR showed the highest bias at
4.11%, indicating significant underestimation. In the test set, LSTM maintained low
bias, confirming its stability across datasets.

Variance (nRMSE): LSTM again showed the lowest variance (2.76% in validation and
4.74% in testing), suggesting high reliability in its forecasts. SVR and MLP had
higher variances (around 14%), indicating lower forecast accuracy for these models.
Performance Summary: LSTM emerged as the most accurate and reliable model, with
minimal bias and variance. Conv-1D and RF performed reasonably well but showed
slightly higher variances. SVR and MLP showed limitations in handling demand
fluctuations accurately.

In summary, LSTM was determined to be the best-performing model for load

forecasting, providing the lowest nMBE and nRMSE values, which underscores its
ability to capture complex temporal patterns in electricity consumption with high
accuracy and stability.

Paper 4
The study “Medium-Term Regional Electricity Load Forecasting through Machine
Learning and Deep Learning” by Shirzadi et al. investigates electricity load
forecasting using machine learning techniques, including Support Vector Machine
(SVM), Random Forest (RF), and deep learning methods like Long Short-Term Memory
(LSTM) and Non-linear Auto-Regressive Exogenous (NARX) neural networks.

Methods:
Data Preparation: The dataset, containing nine years of hourly electricity load and
climatic data (temperature and wind speed) from Bruce County, Ontario, was
preprocessed with feature selection, outlier removal, and normalization.

Machine Learning Models:

Random Forest (RF): An ensemble model used for its interpretability and reliability
in handling non-linear relationships.
Support Vector Machine (SVM): Implemented using support vector regression (SVR)
with a linear kernel to handle time-series data.
Deep Learning Models:

LSTM (Long Short-Term Memory): A recurrent neural network model was employed to
capture long-term dependencies in time-series data.
NARX (Non-linear Auto-Regressive Exogenous Model): Used for its suitability in
time-series forecasting, the NARX model outperformed other methods in capturing
seasonal and peak variations in load data.
Results:
The results indicated that the deep learning models, particularly the NARX network,
provided superior forecasting accuracy. Key findings included:

NARX achieved an R-squared value of 0.96 and RMSE of 5.81 MW, outperforming other
models.
LSTM yielded an R-squared of 0.93 and RMSE of 8.3 MW.
RF and SVM had lower accuracy, with R-squared values of 0.871 and 0.877 and RMSE
values of 11.925 MW and 12.308 MW, respectively.

Paper 5:

Methods:
Forecasting Model: The authors developed a Deep Recurrent Neural Network with Long
Short-Term Memory (DRNN-LSTM) to forecast both residential power load and
photovoltaic (PV) power output. This model included:

Schedule variables like time of day, day of the week, etc.

Weather-related variables like temperature, humidity, wind speed, and radiation.
The DRNN-LSTM model structure had an input layer, two RNN-LSTM layers, two general
perception layers, and an output layer.
Optimization Model: For the optimal load dispatch, a community microgrid model was
created, incorporating residential loads, PV arrays, electric vehicles (EVs), and
an energy storage system (ESS). The optimization aimed to minimize costs and
pollutants, using constraints such as energy balance and capacity constraints for
EVs and ESS. PSO was employed for optimizing the model across three scheduling
scenarios: uncoordinated EV charging, ESS adoption, and coordinated EV charging
with ESS.

Results:
The DRNN-LSTM model outperformed traditional methods (MLP and SVM) with lower RMSE,
MAE, and MAPE, and higher PCC values, showing the accuracy of its forecasts.
In the optimization model, coordinated EV charging and ESS adoption significantly
improved the load balance and reduced daily costs by 8.97%.
Statistical values for forecasting results included an RMSE of 2.987, MAE of 2.365,
and MAPE of 7.43% for load forecasting. PV power forecasting achieved an RMSE of
7.536, MAE of 4.369, and MAPE of 15.87%, with PCC values of 0.925 and 0.961 for
load and PV, respectively.
This approach demonstrated the potential for integrating renewable energy,
improving grid stability, and reducing operation costs in community microgrids.