[Solved] CMSC409 Project 3

You are given hourly energy consumption data from 5:00AM to 8:00PM in 1-hour interval, for 3 days.

Using this data, predict the energy consumption of the 4^th day using a single decision unit (neuron). Note: Create code yourself (no external libraries).

The data is given in 1-hour increments in four text files (please download Pr3_data.zip). The training data for the first 3 days are named train_data_X.txt where X is the date. Use this data for training. In this example, time units are hours and consumption in kW.

Once the decision unit is trained, predict the energy consumption for the 4^th date in 1-hour increments from 5:00AM to 8:00PM. Use the testing data to calculate the total error of the prediction. The testing data is in the file test_data_4.txt. (Note: do not use this data for training, the testing data is only for testing.)

Try out the following three single-neuron architectures:

1. a) b)

c)

Figure 1: Architectures

• What are the inputs and outputs for this neuron (physical meaning)?
• Which activation function is used in the three architectures above? Why?
  1. Hint: it should output continuous real values linearly from INF to +INF, not asymptotically ending in 0 and 1).
• Compare the training and testing total error obtained using the architectures on Figure 1:
  2. Train the decision unit on the data from the first 3 days. Report training total error for each of the three days. Present a graph (original data vs. trained model), similar to the Figure 2.
  3. Predict the energy consumption of the 4^th Calculate the total error of your prediction using the data for the 4^th date. Report testing total error. Present a graph (original data vs. trained model), similar to the Figure 2.

4) Report the number of iterations, the learning rate, and data pre-processing steps you may have chosen (such as normalization of input data). Clearly explain why you selected these values and steps.

Figure 2: An example of the plot showing training data vs. model predictions

Note:

1. While training a neuron, you are essentially using a training algorithm to perform polynomial regression for you (to find appropriate values for coefficients weights in this case).
2. Linear activation function really means summation element.