GRU-based Time-Series Forecasting Model to Predict Household Electric Power Consumption

Problem Statement

Develop a GRU-based time-series forecasting model to predict household electric power consumption using historical electricity usage data and evaluate its performance using MAE and RMSE.

Download Dataset: Household_Power_Consumption.txt

Dataset Description

The dataset contains measurements of electric power consumption in one household, recorded over time.

Attribute	Description
Date	Date of measurement
Time	Time of measurement
Global_active_power	Household global active power in kilowatts (kW)
Global_reactive_power	Household global reactive power in kilowatts
Voltage	Average voltage in volts
Global_intensity	Average current intensity in amperes
Sub_metering_1	Energy consumption mainly from kitchen appliances
Sub_metering_2	Energy consumption mainly from laundry appliances
Sub_metering_3	Energy consumption mainly from water heater and air conditioner

For our GRU practical, the main prediction variable will be:

Global_active_power 

Go to Jupyter Notebook

Dataset contain 278,730 rows × 9 columns.

But for our model, we selected only 1 column: Global_active_power, 

time_step = 60  

Total sequences

The formula is:

$$\text{Total Sequences} = \text{Total Rows} - \text{time\_step}$$

Therefore:

$$278,730 - 60 = \mathbf{278,670}$$

So the model creates 278,670 sequences. 

What is inside each sequence?

Each sequence contains 60 consecutive Global_active_power readings as input and 1 next reading as the target.  

Sequence 1:  Rows 0–59       → Target Row 60

Sequence 2:  Rows 1–60       → Target Row 61

Sequence 3:  Rows 2–61       → Target Row 62

...

Sequence 60: Rows 59–118     → Target Row 119

...

Last Sequence: Rows 278669–278728 → Target Row 278729

So here 

X Shape: (278670, 60) 

y Shape: (278670,)

And after reshaping for GRU: 

X Shape: (278670, 60, 1)

278,670 sequences × 60 time steps per sequence × 1 feature.