Model Card for energy-consumption-anomaly-detection

A Random Forest classifier for detecting anomalies in hourly energy meter readings across 200 commercial buildings, with per-building MinMax scaling and engineered temporal/statistical features.

Model Details

Model Description

This model classifies hourly energy meter readings as Normal or Anomaly, enabling automated monitoring of large-scale building infrastructure without manual review of individual meter logs. It was selected after comparing five candidate models (Transformer Autoencoder, LSTM Autoencoder, XGBoost, LightGBM, Random Forest) and achieved the best precision/recall balance on a highly imbalanced dataset (98.5% normal vs. 1.5% anomaly).

• Developed by: Vikaash17
• Model type: Random Forest Classifier (tabular, binary classification)
• Language(s): N/A (tabular numeric data)
• License: Apache 2.0
• Trained from scratch: Yes (not fine-tuned from a pretrained model)

Model Sources

• GitHub Repository: https://github.com/Vikaash-17/Energy_Consumption_Anomaly_Detection
• Hugging Face Repository: https://huggingface.co/Vikaash17/Energy-Consumption-Anomaly-Detection

Note: The trained model file (~222MB) is hosted here on Hugging Face rather than GitHub due to GitHub's file size limits. The Flask web application code, preprocessing pipeline, and dashboard remain in the GitHub repository.

Uses

Direct Use

Bulk anomaly detection on hourly energy meter readings via CSV upload, for facility managers and energy analysts monitoring multiple buildings. Designed to work with the accompanying Flask web application for interactive prediction and visualization.

Out-of-Scope Use

Not intended for buildings outside the original 200-building training set without validation, non-hourly reading frequencies, or meter types substantially different from those in training. Not designed as a general-purpose anomaly detector for non-energy tabular data.

Bias, Risks, and Limitations

• Trained on 200 specific buildings; generalization to new buildings relies on a global fallback scaler and has not been extensively validated.
• Severe class imbalance (1.5% anomalies) means the model may still under-detect rare/novel anomaly patterns not represented in training data.
• Per-building performance varies (F1 range: 0.76–1.00 across buildings) — lower-performing buildings should be monitored more closely.

Recommendations

Users should validate performance on new buildings before production use, and treat predictions for buildings with historically lower F1 scores (e.g., Building 240) with additional scrutiny.

How to Get Started with the Model

⚠️ Note: This model is serialized using Python's pickle module. Only load pickle files from sources you trust, as they can execute arbitrary code during deserialization.

import pickle
import pandas as pd
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="Vikaash17/Energy-Consumption-Anomaly-Detection",
    filename="model.pkl"  # replace with your actual filename in the repo
)

with open(model_path, "rb") as f:
    model = pickle.load(f)

# Input requires 15 engineered features (see Training Data section)
# X = preprocessed_dataframe[feature_columns]
predictions = model.predict_proba(X)[:, 1]
anomaly_flag = predictions >= 0.20  # optimal threshold

Training Details

Training Data

Hourly energy meter readings from 200 commercial buildings over a full year (8,736 hours), ~1.4M+ records. Raw inputs: building_id, timestamp, meter_reading. Anomaly labels were generated via statistical deviation analysis. Dataset is not publicly available.

Feature Engineering

15 engineered features computed from the 3 raw inputs, requiring a minimum of 49 prior readings per building:

Feature	Description
meter_reading_scaled	Per-building MinMax scaled reading (scaler fit on normal data only)
lag_1, lag_2, lag_3	Readings 1/2/3 hours prior
lag_24, lag_48	Readings 24/48 hours prior
rolling_mean_24, rolling_std_24, rolling_max_24	24-hour rolling statistics
hour_sin, hour_cos	Cyclical hour-of-day encoding
day_sin, day_cos	Cyclical day-of-month encoding
month_sin, month_cos	Cyclical month encoding

Training Procedure

• Train/test split: 80% / 20%
• Class-imbalance-aware evaluation (98.5% normal / 1.5% anomaly)
• Detection threshold tuned via precision-recall curve analysis (optimal: 0.20, vs. default 0.50)

Evaluation

Model Comparison

Model	Precision	Recall	F1 (Anomaly)	Macro F1
Transformer Autoencoder	0.30	0.02	0.04	0.50
LSTM Autoencoder	0.19	0.28	0.23	0.58
XGBoost	0.51	0.89	0.65	0.76
LightGBM	0.49	0.90	0.64	0.75
Random Forest (this model)	0.81	0.89	0.85	0.92

Global Classification Report (Threshold = 0.20)

Class	Precision	Recall	F1 Score	Support
Normal (0)	1.00	1.00	1.00	341,304
Anomaly (1)	0.81	0.89	0.85	6,675
Accuracy			0.99	347,979
Macro Avg	0.90	0.94	0.92	347,979
Weighted Avg	0.99	0.99	0.99	347,979

Threshold Optimization

Threshold	Precision	Recall	F1 Score
0.50 (default)	0.97	0.68	0.80
0.40	0.96	0.72	0.82
0.30	0.94	0.76	0.84
0.20 (optimal)	0.81	0.89	0.85
0.15	0.84	0.83	0.83

Per-Building Performance

Metric	Value
Mean F1 across 200 buildings	0.96
Buildings with F1 = 1.00	30+
Buildings with F1 ≥ 0.90	~185 / 200
Buildings with F1 < 0.85	10 / 200
Minimum F1 (Building 240)	0.76

Summary

Random Forest outperformed all other candidates, including deep learning approaches, largely due to the tabular nature of the feature set and severe class imbalance, which disadvantaged the autoencoder-based models. XGBoost and LightGBM had strong recall but weaker precision (more false alarms). Threshold tuning improved anomaly recall from 0.68 to 0.89 vs. the default 0.50 threshold.

Technical Specifications

Model Architecture and Objective

Random Forest Classifier (scikit-learn) with tuned decision threshold (0.20), trained for binary anomaly classification on engineered tabular features.

Compute Infrastructure

Software

• Python 3.x
• scikit-learn
• Pandas, NumPy
• Pickle (model serialization)
• Flask (deployment/web app)

Model Card Contact

Vikaash17 — via Hugging Face profile or GitHub repository issues.