🌊 Time Series Water Quality Forecasting

---

📌 Overview

This project applies time series analysis and forecasting to predict Dissolved Oxygen (O₂) concentrations in the Southern Bug (Pivdennyi Booh) River, Ukraine, using 21+ years of multivariate water quality monitoring data.

Dissolved Oxygen is one of the most vital indicators of river health — it directly reflects biological activity, water temperature dynamics, and ecosystem sustainability. This project builds a full forecasting pipeline from raw data ingestion through model comparison and future projection.

Goal: Predict O₂ at time t using past measurements, capturing seasonal and trend dynamics through both classical statistical models and machine learning.

---

📊 Dataset

Property	Detail
Source	Kaggle — Southern Bug River Water Quality
Stations	22 monitoring stations
Records	2,861 raw observations
Period	January 2000 – April 2021
Focus Station	Station 3 — 262 observations over 21.1 years
Target Variable	Dissolved Oxygen (O₂) in mg/L
Features	NH4, BSK5, Suspended, NO3, NO2, SO4, PO4, CL

---

🔬 Project Pipeline

Raw Data (Kaggle)
      │
      ▼
Data Preparation
  ├── Date parsing & sorting
  ├── Station filtering (Station 3)
  ├── Missing value imputation (median)
  └── Monthly resampling + interpolation (254 months)
      │
      ▼
Exploratory Data Analysis
  ├── Descriptive statistics
  ├── Time series visualization (all 9 parameters)
  ├── Seasonal box plots
  ├── Correlation heatmap
  └── Additive decomposition (trend + seasonality + residual)
      │
      ▼
Stationarity Testing
  ├── ADF Test  →  Original: NON-STATIONARY | Differenced: STATIONARY
  └── KPSS Test →  Original: NON-STATIONARY | Differenced: STATIONARY
      │
      ▼
ACF & PACF Analysis
  └── Seasonal spikes at lags 6, 12, 18, 24 → confirms s=12
      │
      ▼
Model Building (Train: 230m | Test: 24m)
  ├── ARIMA(1,1,1)
  ├── SARIMA(1,1,1)(1,1,1,12)
  ├── Holt-Winters Exponential Smoothing
  └── Gradient Boosting + Lag Feature Engineering
      │
      ▼
Model Evaluation & Selection
      │
      ▼
24-Month Forecast (2021–2023)

---

🧠 Models Used

Model	Type	Seasonal Aware
ARIMA(1,1,1)	Statistical	❌
SARIMA(1,1,1)(1,1,1,12)	Statistical	✅
Holt-Winters Exponential Smoothing	Statistical	✅
Gradient Boosting + Lag Features	Machine Learning	✅ (via lag_12, month_sin/cos)

---

📈 Results

Model Comparison Table

Rank	Model	MAE	RMSE	MAPE
🥇 1	Holt-Winters	1.037	1.417	9.01%
🥈 2	SARIMA(1,1,1)(1,1,1,12)	1.103	1.491	9.39%
🥉 3	Gradient Boosting	1.098	1.466	9.65%
4	ARIMA(1,1,1)	3.351	4.096	37.83%

✅ Best Model: Holt-Winters with MAPE = 9.01% and MAE = 1.037 mg/L

The three seasonal-aware models all performed within a narrow band (~9% MAPE), confirming that capturing the annual cycle is the single most critical factor for accurate O₂ forecasting. The non-seasonal ARIMA baseline failed significantly, underscoring the importance of the seasonal component.

---

🔭 24-Month Forecast (2021–2023)

Using the best model (Holt-Winters), a 24-month forward forecast was produced with 95% confidence intervals:

Date	Forecast (mg/L)	Lower 95% CI	Upper 95% CI
2021-03	13.840	10.104	17.576
2021-07 (summer min)	7.893	3.939	11.847
2022-01 (winter peak)	14.041	9.974	18.109
2022-07 (summer min)	7.892	3.666	12.118
2023-01 (winter peak)	14.053	9.705	18.401
2023-02	14.330	9.962	18.698

O₂ levels are forecast to remain within environmentally safe ranges, with summer minima staying above ~7.9 mg/L through 2023.

---

🧪 Key Findings

• Strong Seasonality — Seasonal strength score of 0.697 (Strong), with O₂ peaking in winter (cold water holds more oxygen) and dipping in summer
• Moderate Downward Trend — Trend strength of 0.314 (Moderate), consistent with gradual surface water warming over the 21-year period
• Stationarity — The original series is non-stationary (ADF p = 0.29); first differencing achieves stationarity (ADF p ≈ 0.00)
• ACF/PACF — Clear annual seasonality at lag 12, motivating SARIMA with s=12
• Seasonality is Critical — The gap between ARIMA (37.83% MAPE) and SARIMA (9.39% MAPE) confirms that ignoring seasonality renders forecasts unreliable

---

🛠️ Tech Stack

pandas        # Data manipulation & resampling
numpy         # Numerical operations
matplotlib    # Time series & decomposition plots
seaborn       # Correlation heatmap & box plots
statsmodels   # ARIMA, SARIMA, Holt-Winters, ADF/KPSS, ACF/PACF
scikit-learn  # Gradient Boosting, imputation, metrics
scipy         # Statistical utilities
kagglehub     # Dataset download

Environment: Google Colab with TPU v5e accelerator

---

🚀 Getting Started

1. Clone the Repository

git clone https://github.com/MuhammadUsman-Khan/Time-Series-Water-Quality-Forecasting.git
cd Time-Series-Water-Quality-Forecasting

2. Install Dependencies

pip install pandas numpy matplotlib seaborn statsmodels scikit-learn scipy kagglehub

3. Run the Notebook

Open Time_Series_Water_Quality_Forecasting.ipynb in Jupyter or Google Colab.

The dataset is automatically downloaded from Kaggle via kagglehub — no manual download needed.

import kagglehub
path = kagglehub.dataset_download("vbmokin/wq-southern-bug-river-01052021")

⚠️ You will need a Kaggle account and API credentials configured for kagglehub to work.

---

📁 Repository Structure

Time-Series-Water-Quality-Forecasting/
│
├── Time_Series_Water_Quality_Forecasting.ipynb   # Main notebook
└── README.md                                      # Project documentation

---

🔮 Future Improvements

• Extend analysis to all 22 monitoring stations for spatial comparison
• Incorporate exogenous variables (water temperature, flow rate) via SARIMAX
• Evaluate deep learning models (LSTM, Temporal Convolutional Networks)
• Build an ensemble forecasting framework to reduce confidence interval width
• Deploy a lightweight forecasting dashboard using Streamlit

---

🙋‍♂️ Author

Muhammad Usman Khan

---

If you found this project useful, consider giving it a ⭐ — it helps others discover it!