Engineering · Machine learning

Predictive CFD with Neural Networks

A building-simulation study predicting indoor temperature from simulation inputs using Python, a feed-forward neural network, ADAM optimization and held-out test comparisons.

The test compares predicted indoor temperatures with the held-out source simulation outputs.
Diagram in text
Source simulations
More than 20,000 hourly records.
Prepare inputs
Normalize 13 inputs and split the data 70/30.
Train the network
Fit a 13 → 13 → 8 → 1 neural network.
Evaluate predictions
Compare held-out temperatures with simulation.

Approach and contribution

30-second case brief

A 13-input model tested against held-out simulation records.

Problem
Explore a compact surrogate for indoor temperature using the operating conditions and surface temperatures from building simulations.
Contribution
Prepared data, developed the neural-network model, evaluated predictions, and contributed to the technical report within a three-person team.
Result
Trained a 13 → 13 → 8 → 1 network using more than 20,000 hourly records and compared its temperature predictions with held-out source-simulation outputs.

Approach70/30 train-test split · normalized inputs · ADAM · 100 epochs · prediction/reference comparison.

Project scope and responsibility

Context
Drexel University
Role
Data preparation, ANN development, evaluation, and technical reporting within a three-person team
Dataset
20,000+ hourly observations · 13 inputs
Model
Feed-forward ANN · ADAM optimization
Team
V. Edske · J. Moussa · R. Tajik

Result evidence

Team final report: reference simulation and ANN temperature predictions plotted across held-out observations.

Overview

Detailed building simulations can be slow to rerun during iterative analysis. This project tested whether an artificial neural network could learn the relationship between 13 simulation inputs and indoor ambient temperature, creating a faster surrogate for a defined case and dataset.

A compact architecture was trained on normalized simulation data.

Evaluation

More than 20,000 hourly records were separated into 70 percent training and 30 percent testing subsets. The selected configuration used sigmoid activation, ADAM optimization, a batch size of 36, and 100 epochs. Error definitions and validation scope remain visible alongside the result.

Correlation checks complement the summary error metric.

Engineering record

Technical dossier

Can a compact neural network predict occupant-level indoor temperature from building simulation inputs?

Contribution
Data preparation, ANN development, evaluation, visual analysis, and technical reporting in a three-person team.
System boundary
More than 20,000 hourly records, 13 predictors, and one indoor ambient-temperature output.

Method and system architecture

  1. 01

    Define

    Separate the 13 predictors from the indoor-temperature target.

  2. 02

    Prepare

    Inspect complete records, standardize variables, and define training/testing subsets.

  3. 03

    Train

    Evaluate batch sizes and epoch counts for the dense neural network.

  4. 04

    Evaluate

    Compare predicted and reference temperatures in the held-out set and selected test observations.

Thirteen input features

Feature groupCountVariables
Occupant context3Number of occupants; x and y coordinates
Surface temperatures8Four walls, two windows, floor, ceiling
Supply air2Mass flow rate and temperature

Source: Final report p. 6, Figure 3.

Network and training configuration

SettingDocumented value
Layer sizes13 inputs → 13 hidden → 8 hidden → 1 output
Hidden activationSigmoid
OutputDense linear regression output in archived implementation
OptimizerADAM
Training lossMean squared error in archived implementation
Batch size / epochs36 / 100
Train / test70% / 30%; random_state = 42
Evaluation viewsPrediction correlation, sample comparison, selected test observations

Source: Final report pp. 6-10 and archived team implementation.

Standardization

z = (x - mean) / standard deviation

Dense layer

h = sigmoid(Wx + b)

Regression training loss

MSE = mean((predicted temperature - reference temperature)²)

Verification and quality checks +
Shape and schema
Connect each feature column to the 13-input model and the single target.
Test separation
Compare predictions with reference values in the recorded held-out subset.
Error inspection
Read scatter and sample plots together to locate temperature-dependent prediction behavior.

The complete project record

Full book

Read the complete report, thesis, or presentation here. Use the page controls to explore the original figures and methods.

Download PDF Open full PDF

11 pages · V. Edske, J. Moussa, R. Tajik · December 2022

Artificial Neural Network for Predictive CFD

The original document, page by page.

The reader loads when you open the book.

Contact

Do you have a question or want to chat?

Analytics preferences

Optional Google Analytics uses cookies to measure page visits and engagement. You can browse the entire portfolio with analytics off.

Analytics is off in this browser.

Privacy details. Change your choice here at any time. Your preference is remembered for up to 180 days.