[Paper Review] MLRM: A Multiple Linear Regression based Model for Average Temperature Prediction of A Day

Ishu Gupta, Harsh Mittal+2 more|arXiv (Cornell University)|Mar 11, 2022

Hydrological Forecasting Using AI21 citations

TL;DR

The paper presents a multiple linear regression model with feature selection to predict a day's mean temperature from the previous three days, achieving about 2.8°C error on the test set.

ABSTRACT

Weather is a phenomenon that affects everything and everyone around us on a daily basis. Weather prediction has been an important point of study for decades as researchers have tried to predict the weather and climatic changes using traditional meteorological techniques. With the advent of modern technologies and computing power, we can do so with the help of machine learning techniques. We aim to predict the weather of an area using past meteorological data and features using the Multiple Linear Regression Model. The performance of the model is evaluated and a conclusion is drawn. The model is successfully able to predict the average temperature of a day with an error of 2.8 degrees Celsius.

Motivation & Objective

• Predict the mean temperature of a day using weather data from the prior three days.
• Apply feature selection to remove non-linear and insignificant features before regression.
• Evaluate model performance with a train/test split and visualize predicted versus actual temperatures.
• Demonstrate feasibility with Weather Underground data and assess predictive accuracy.

Proposed method

• Compute Pearson correlation to assess linear relationships and select features with |r| > 0.6.
• Apply backward elimination using p-values at a 0.05 significance level to discard insignificant features.
• Split data into 80% training and 20% testing for the multiple linear regression model.
• Train the MLR model to minimize the sum of squared errors between actual and predicted temperatures.
• Use Python libraries (Pandas, StatsModels, scikit-learn, matplotlib) for feature selection, model training, evaluation, and visualization.

Experimental results

Research questions

• RQ1 What is the achievable mean absolute error/mean squared error in predicting daily mean temperature using the proposed MLR approach?
• RQ2 Which features among the initial set remain significant after correlation-based screening and backward elimination?
• RQ3 How does the MLR model perform on unseen data when predicting the mean temperature of a day from the past three days?
• RQ4 What is the role of feature selection in improving the linear regression model for this weather prediction task?

Key findings

• The study used a Weather Underground dataset with 997 instances.
• After feature selection, 7 features remained for the MLR model.
• The model achieved an absolute mean error of 2.8 degrees Celsius on the test set.
• Pearson correlation filtering ensured all remaining features had |r| > 0.6 before backward elimination.
• The final model’s predictions, when plotted, align closely with the y = x line, indicating good accuracy.