Flex›Source - Load Forecasting Methods Survey (2025)

Source - Load Forecasting Methods Survey (2025)

Hasan et al., “A comprehensive review of load forecasting techniques for smart grids,” Energy Conversion and Management: X 26 (2025) 100922. Bangladeshi/UK academic group (Mahmudul Hasan et al., eight authors). 121 references. Published 2025 open-access. Covers all four forecasting horizons with comparative method analysis, performance metrics, and smart grid integration context.

Relevance to wiki

Fills the operational forecasting layer between strategic/regulatory content and actual market participation. STLF (day-ahead to 1 week) is the data layer required for BRP plan submission, FSP bidding, and explicit demand response — a dimension not covered by the Energiforsk sources on DNDP/FNA (which use MTLF/LTLF). The user established the relevance: “the connection to flex markets is that you need short-term load forecasting to do anything operationally (unknown demand = no trading).”

Forecasting horizon taxonomy

Four horizons defined and compared:

HorizonTimeframePrimary use
VSTLFMinutes to hoursReal-time grid control, FCR/aFRR dispatch
STLFDay-ahead to 1 weekBRP plan submission, FSP bidding, DR scheduling
MTLF1 week to 1 yearMaintenance planning, capacity optimization
LTLF> 1 yearNetwork development planning, DNDP, FNA

Methods landscape

Statistical methods:

  • ARIMA/SARIMA — time-series models; SARIMA adds seasonal component; ARIMAX adds exogenous variables (weather, time-of-day); good baseline; limited non-linear capture
  • Multiple regression — interpretable; good for MTLF when relationships are stable
  • Exponential smoothing — simple; computationally cheap; weak on structural breaks

Machine learning:

  • ANN (Artificial Neural Network) — flexible non-linear fitting; training-data intensive; black-box
  • SVM/SVR (Support Vector Regression) — good for limited data; kernel choice matters
  • Random Forest, XGBoost — ensemble methods; strong feature importance; widely used for STLF

Deep learning:

  • LSTM — captures long-term temporal dependencies; most studied for STLF; sequential data specialist
  • CNN — extracts local temporal patterns; faster training than RNN variants
  • BiLSTM — bidirectional LSTM; captures forward and backward temporal context
  • GRU — lighter LSTM variant; comparable performance, lower compute
  • Transformer — attention mechanism; captures non-sequential dependencies

Hybrid models — consistently best results:

CNN-LSTM, CNN-GRU, CNN-BiLSTM-Bayesian, and LSTM-Transformer all outperform their components. Quantified comparison from the paper: LSTM-BPNN achieved 1.47% MAPE vs 3.55% (LSTM alone) vs 4.09% (BPNN alone). CNN-BiLSTM with Bayesian optimization resolves gradient explosion/disappearance issues that afflict pure LSTM on longer sequences.

The paper’s overall conclusion: “ML based technique used for load forecasting produces more precise, reliable and accurate predictions than statistical approaches. Studies showed that combined prediction model or hybrid model presented better results.”

Key input variables

Temperature is the dominant driver at all timescales. Secondary inputs: historical load, time features (hour-of-day, day-of-week, season, holidays), economic activity indicators for MTLF/LTLF. EV penetration and DR participation are increasingly important as both reshape load patterns and invalidate historical training data.

Smart meter data effects

Granular AMI data enables individual and building-level STLF, supporting real-time disaggregation and DR dispatch. Data quality (noise, missing values) degrades accuracy; low-voltage and microgrid forecasting has higher relative variability than bulk-system forecasting.

EV and DR integration challenge

EVs create new load patterns not represented in historical training data. DR participation adds demand-side variability. Both make STLF harder as penetration grows — but also more valuable, since unknown demand makes market bidding impossible.

Performance metrics

MAPE (Mean Absolute Percentage Error) is the most commonly reported metric across comparative studies. RMSE (Root Mean Square Error) penalizes large errors more heavily than MAE (Mean Absolute Error). R² measures explained variance; MSE is RMSE squared. Hybrid models consistently show lower MAPE than single-method models across the reviewed literature.

Future directions

Paper flags: novel pre-processing for power/wind uncertainty; additional topographic and climatic inputs via IoT sensors; improved non-linear fitting by mixing models; integrated wind/solar/load forecasting as a joint problem (predicting renewable output first, then using as input for load forecast).