Smart Building Energy Optimization

This portfolio project builds an end-to-end energy analytics workflow for a smart company building: source ingestion, PostgreSQL normalization, hourly energy-balance reconstruction, load forecasting, BESS simulation, and experiment reporting.

Load forecasting compares seasonal-naive baselines with histogram gradient boosting under chronological evaluation. Battery dispatch is compared using a transparent heuristic and a 24-hour rolling-horizon linear program (LP). The project is intended for learning and technical demonstration, not as production EMS software.

What This Project Demonstrates

• ingestion and normalization of heterogeneous energy time series
• PostgreSQL schema and analysis-view design
• leakage-safe 24-hour load forecasting with chronological evaluation
• physically validated battery dispatch simulation
• rule-based control compared with mathematical optimization
• reproducible, parallelized capacity-sensitivity experiments

Key Results

Battery Optimization

• The site already self-consumes about 92% of local PV and CHP generation.
• The best tested 1000 kWh case saves about 13.4k EUR/year in simulated operating cost.
• In dynamic grid-charging operation, the LP adds about 4.0k EUR/year over the heuristic by scheduling energy for more valuable discharge hours.
• Larger batteries increase total savings and surplus capture, but show diminishing marginal value and fewer equivalent cycles per installed kWh.

Savings exclude BESS purchase, installation, financing, maintenance, demand charges, and replacement costs. Savings are measured against the corresponding no-battery baseline: fixed-price scenarios use the fixed-price baseline, while dynamic-price scenarios use the dynamic-price baseline. See the full experiment results for capacity sensitivity, utilization, runtime, feasibility checks, and a 48-hour dispatch comparison.

Load Forecasting

The fixed HGB model achieves 8.21% WAPE on the held-out 2021 test period, with performance close to its Q4 2020 validation result. Error analysis shows that summer working hours remain the clearest opportunity for improvement.

See the load forecast results for the chronological evaluation design, model comparison, and error diagnostics.

System Overview

Dryad building measurements
    -> PostgreSQL normalization and hourly energy reconstruction
       -> load forecasting
          -> chronological validation and frozen 2021 test
       -> battery simulation + SMARD prices
          -> heuristic and rolling-horizon LP dispatch
          -> physical validation and experiment metrics

Data Sources

The building source is the corrected reduced_data.zip version updated on February 26, 2025. Load forecasting uses hourly electricity measurements from June 2019 through 2021; the BESS experiments use the 2021 measurements. The local archive size, approximately 320.16 MB, matches that corrected Dryad release.

Engel, Jens; Castellani, Andrea; Wollstadt, Patricia et al. (2025). A real-world energy management data set from a smart company building for optimization and machine learning [Dataset]. Dryad. https://doi.org/10.5061/dryad.73n5tb363

Dryad datasets are published under CC0; the citation is retained to credit the dataset authors. German day-ahead electricity prices are sourced from SMARD, operated by the German Federal Network Agency.

Quick Start

Download reduced_data.zip from the Dryad dataset and place it at data/reduced_data.zip. Then run:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
cp .env.example .env
docker compose up -d
python scripts/ingest_data.py

Run Battery Experiments

# Hourly capacity-sensitivity experiment
python scripts/battery/run_capacity_analysis.py

# General BESS simulation (15-minute resolution by default)
python scripts/battery/run_bess_simulation.py

# Compare hourly and 15-minute resolution
python scripts/battery/run_bess_simulation.py --resolutions hour 15min

Shared assumptions are defined in scripts/battery/experiment_defaults.py. Configuration options, terminal-value comparisons, and audit exports are documented in the BESS LP methodology.

Run Load Forecasting

python scripts/forecasting/run_validation.py
python scripts/forecasting/run_final_test.py

Both commands write compact CSV summaries to results/forecasting/. Pass --save-forecasts only when the large row-level prediction and error table is needed. The final-test command reproduces the fixed 2021 result.

requirements.txt contains the runtime dependencies. requirements-dev.txt adds test, notebook, and figure-generation tooling for local development.

To run the offline unit tests, install the development dependencies and run pytest:

python -m pip install -r requirements-dev.txt
python -m pytest tests -q

The unit tests are offline: they do not require the downloaded source archive, network access, or PostgreSQL.

Documentation

Document	Contents
Battery experiment results	Findings, charts, capacity sensitivity, runtime, and limitations
BESS time-resolution comparison	Hourly versus 15-minute energy, peaks, and simulated operating value
BESS terminal-value comparison	Rolling-horizon A/B test with and without terminal SOC valuation
Battery simulation methodology	Energy conventions, pricing, battery model, metrics, and validation
Heuristic dispatch	Rule-based controller and rolling price thresholds
LP optimization	Objective, constraints, rolling horizon, and modeling choices
Load forecast results	Method, validation, final 2021 results, and error diagnostics

Detailed assumptions, scope boundaries, and limitations are documented with the corresponding methodology and results.