Hyperparameter Optimization of Regression Model for Electrical Load Forecasting During the COVID-19 Pandemic Lockdown Period

Q: How does K-nearest neighbors (KNN) algorithm make predictions?

The K-nearest neighbors (KNN) algorithm makes predictions by utilizing a similarity measure, specifically a distance function. It assigns a value to a new data sample based on its proximity to the training set points. The algorithm uses a metric such as Euclidean, Manhattan, or Minkowski distance to calculate the distance between each sample in the test set and every sample in the training set. The K nearest samples of training data are then picked. For regression, the KNN prediction is the mean of the outcomes from the K nearest neighbors. This approach allows the algorithm to make predictions for new data points based on their similarity to the training set.

Q: What are the essential components of hyper-parameter optimization?

The essential components of hyper-parameter optimization are an estimator with its objective function, a search space (configuration space), a search or optimization method, and an evaluation function. The objective function aims to minimize the root mean squared error (RMSE) or error rate. The search space represents the possible values for hyper-parameters, and the optimization method, such as Particle Swarm Optimization (PSO) or Genetic Algorithms (GA), locates the optimal hyper-parameter combinations. The evaluation function compares the performance of various hyper-parameter configurations to identify the best practicable results.

Q: How were hyperparameters optimized in the experiments?

In the experiments, optimization algorithms such as Particle Swarm Optimization (PSO) and Genetic Algorithms (GA) were used to optimize the hyperparameters of machine learning models. The initial hyperparameters values for PSO and GA algorithms were provided in Tables 5 and 6, based on the default settings of the Python scikit-learn library package. The best hyperparameters obtained with different optimization techniques were shown in Tables 7 and 8. The optimization process involved 5-fold cross-validation to establish the optimal model parameters for Sacramento, Los Angeles, and New York City. Convergence was reached before 100 iterations, and the process was stopped when the variance between iterations was less than 0.5%.

Q: How does KNN-PSO model perform in Sacramento and New York?

The KNN model performs the best in Sacramento with an RMSE value of 3.96, but performs the worst in New York with an RMSE value of 19.65. The KNN regressor is easier to implement due to its lower number of hyperparameters (one) compared to the SVM regressor (six) and RF regressor (three). The KNN-PSO model outperforms other models in terms of accuracy and is more resilient in coping with missing data. It was chosen to assess the effect of COVID-19 on city-scale power consumption due to its superior accuracy compared to other models. The model was trained with data from July 2015 to March 2020 and used to forecast power consumption after the pandemic. The results showed a disparity between expected and actual electricity usage after the lockdown period, indicating that the COVID-19 pandemic altered electricity demand patterns.

Q: How did machine learning models perform in forecasting power consumption?

The study utilized RF, SVM, and KNN models to forecast power consumption in New York, Los Angeles, and Sacramento. Hyperparameter tuning improved the accuracy of predictions. SVM, RF, and KNN outperformed PSO and GA metaheuristics. KNN had the highest accuracy, while PSO had the lowest RMSE score. The ARIMA model was also compared. During the COVID-19 lockdown, power consumption declined by 2-12% compared to 2019.

Question

1. How does predicting electricity demand at the city level contribute to urban sustainability and energy efficiency?

2. How was the primary dataset divided for Phase 1: Data preparation?

3. What is the ARIMA model structure?

4. What is the error term penalty parameter in SVM models?

Accepted Answer

Predicting electricity demand at the city level is crucial for promoting urban sustainability and increasing global energy efficiency. It helps in power plant energy efficiency, resource planning, greenhouse gas emissions tracking, program evaluation, reserve requirements, and system infrastructure analysis. By understanding building energy consumption on a city-wide scale, policymakers and energy planners can develop effective climate change solutions and improve the preparedness of energy production and transmission infrastructure for high-demand situations. This, in turn, enhances resilience and energy security in the context of climate change adaptation. Additionally, predicting electricity demand based on factors such as temperature, population, and income can aid in reducing carbon emissions and optimizing energy usage in cities.

Accepted Answer

The primary dataset was divided into power data and weather data. Power data included date, time, electricity load, and day type (working day/non-working day). Weather data included date, time, ambient temperature, and humidity. These datasets were merged to prepare for study phases. A new feature was added, comprising daily electricity load, demand peaks, and average ambient temperature. Missing values were eliminated, and the dataset was explored for variable scales. Finally, 80% of the dataset was used for model training, and 20% for validation testing.

Accepted Answer

The ARIMA model structure consists of three components: autoregressive (AR), moving average (MA), and differencing (I). The AR component represents the relationship between an observation and a certain number of lagged observations. The MA component represents the relationship between an observation and the residual errors from a moving average model applied to lagged observations. The differencing component is used to make the time series stationary by removing trends and seasonality. The ARIMA model is denoted as ARIMA(p, d, q), where p is the order of the AR component, d is the order of differencing, and q is the order of the MA component. The model is fitted by estimating the parameters using the partial auto-correlation function (PACF) and autocorrelation function (ACF) of the time series data.

Accepted Answer

The error term penalty parameter, denoted as C, is a critical hyper-parameter in all SVM models. It represents the normalization vector and plays a crucial role in achieving linear separability. By adjusting the penalty parameter, the SVM algorithm can control the trade-off between maximizing the margin and minimizing the classification error. A higher penalty parameter results in a smaller margin but a lower classification error, while a lower penalty parameter leads to a larger margin but a higher classification error. Therefore, selecting an appropriate value for the penalty parameter is essential for achieving optimal performance in SVM models.

Accepted Answer

The mtry parameter in the Random Forest (RF) algorithm specifies the number of variables used for node splitting during tree creation. It is set to a value between 1 and n, where n is the total number of variables. The purpose of choosing mtry < n is to reduce computational time and complexity. By limiting the number of variables considered at each split, the algorithm can build trees more efficiently. This parameter helps in controlling the size and depth of the trees, preventing overfitting and improving the overall performance of the RF model. Additionally, mtry plays a crucial role in the diversity of the trees in the forest, as different trees will consider different subsets of variables, leading to a more robust and accurate prediction.

Accepted Answer

The K-nearest neighbors (KNN) algorithm makes predictions by utilizing a similarity measure, specifically a distance function. It assigns a value to a new data sample based on its proximity to the training set points. The algorithm uses a metric such as Euclidean, Manhattan, or Minkowski distance to calculate the distance between each sample in the test set and every sample in the training set. The K nearest samples of training data are then picked. For regression, the KNN prediction is the mean of the outcomes from the K nearest neighbors. This approach allows the algorithm to make predictions for new data points based on their similarity to the training set.

Accepted Answer

The essential components of hyper-parameter optimization are an estimator with its objective function, a search space (configuration space), a search or optimization method, and an evaluation function. The objective function aims to minimize the root mean squared error (RMSE) or error rate. The search space represents the possible values for hyper-parameters, and the optimization method, such as Particle Swarm Optimization (PSO) or Genetic Algorithms (GA), locates the optimal hyper-parameter combinations. The evaluation function compares the performance of various hyper-parameter configurations to identify the best practicable results.

Accepted Answer

In the experiments, optimization algorithms such as Particle Swarm Optimization (PSO) and Genetic Algorithms (GA) were used to optimize the hyperparameters of machine learning models. The initial hyperparameters values for PSO and GA algorithms were provided in Tables 5 and 6, based on the default settings of the Python scikit-learn library package. The best hyperparameters obtained with different optimization techniques were shown in Tables 7 and 8. The optimization process involved 5-fold cross-validation to establish the optimal model parameters for Sacramento, Los Angeles, and New York City. Convergence was reached before 100 iterations, and the process was stopped when the variance between iterations was less than 0.5%.

Accepted Answer

The KNN model performs the best in Sacramento with an RMSE value of 3.96, but performs the worst in New York with an RMSE value of 19.65. The KNN regressor is easier to implement due to its lower number of hyperparameters (one) compared to the SVM regressor (six) and RF regressor (three). The KNN-PSO model outperforms other models in terms of accuracy and is more resilient in coping with missing data. It was chosen to assess the effect of COVID-19 on city-scale power consumption due to its superior accuracy compared to other models. The model was trained with data from July 2015 to March 2020 and used to forecast power consumption after the pandemic. The results showed a disparity between expected and actual electricity usage after the lockdown period, indicating that the COVID-19 pandemic altered electricity demand patterns.

Accepted Answer

The study utilized RF, SVM, and KNN models to forecast power consumption in New York, Los Angeles, and Sacramento. Hyperparameter tuning improved the accuracy of predictions. SVM, RF, and KNN outperformed PSO and GA metaheuristics. KNN had the highest accuracy, while PSO had the lowest RMSE score. The ARIMA model was also compared. During the COVID-19 lockdown, power consumption declined by 2-12% compared to 2019.

Hyperparameter Optimization of Regression Model for Electrical Load Forecasting During the COVID-19 Pandemic Lockdown Period

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. How does predicting electricity demand at the city level contribute to urban sustainability and energy efficiency?

2. How was the primary dataset divided for Phase 1: Data preparation?

3. What is the ARIMA model structure?

4. What is the error term penalty parameter in SVM models?

5. What is the role of mtry in RF algorithm?

6. How does K-nearest neighbors (KNN) algorithm make predictions?

7. What are the essential components of hyper-parameter optimization?

8. How were hyperparameters optimized in the experiments?

9. How does KNN-PSO model perform in Sacramento and New York?

10. How did machine learning models perform in forecasting power consumption?

Citations

Hyperparameter Optimization of Regression Model for Electrical Load Forecasting During the COVID-19 Pandemic Lockdown Period

Detection and Diagnosis of Different Types of Obesity Using Machine Learning-Based Approaches

Analysis of Electrical Power Losses in Low-Voltage Distribution Networks: A Study of Technical and Non-Technical Losses

A modified grid search-based optimization for possibly repetitive global extremum with an application to edge intelligence in IIoT towards time-domain signals

Pioneering Advanced Deep Multi-Task Learning Model for Precise COVID-19 Diagnosis

References

Climate Change, Mortality, and Adaptation: Evidence from Annual Fluctuations in Weather in the US

Forecasting electricity consumption: A comparison of regression analysis, neural networks and least squares support vector machines

Evolutionary Resilience and Strategies for Climate Adaptation

Deep learning framework to forecast electricity demand

Electric load forecasting in smart grids using Long-Short-Term-Memory based Recurrent Neural Network

Related Papers (5)

Suggestions from Cremona, Italy: 2 months into the pandemic at the frontline of COVID-19 in Europe.

Q&A with Dean Winslow, MD, on Volunteering in Antarctica Currently the only continent without confirmed cases of COVID-19 during COVID-19 Pandemic.

Increased Mortality in Germany and in the Individual German States During the SARS-CoV-2/COVID-19 Pandemic in 2020 and 2021.

PP018 [Infections » Covid-19 / Sars-CoV-2]: COMPARISON BETWEEN FIRST AND SECOND WAVE OF SARS-COV-2 PANDEMIC: CLINICAL PROFILES AND PREDICTORS OF SEVERE COVID-19 CHILDREN

A 10-step guide to convert a surgical unit into a COVID-19 unit during the COVID-19 pandemic.