Short-Term Forecasting of Energy Production and Consumption Using Extreme Learning Machine: A Comprehensive MIMO based ELM Approach

This paper proposes a novel Multi-Input Multi-Output (MIMO) Extreme Learning Machine (ELM) approach for short-term energy forecasting in Corsica that leverages sliding windows and cyclic time encoding to achieve high accuracy and computational efficiency for real-time applications, significantly outperforming persistence models while offering a closed-form solution superior to deep learning methods like LSTM.

The Gist

Imagine you are the conductor of a massive, chaotic orchestra. This isn't a symphony of violins and flutes, but an orchestra of energy: solar panels, wind turbines, hydroelectric dams, diesel generators, and electricity imported from the mainland. Your job is to keep the music playing perfectly, ensuring there is just enough power for everyone without wasting a single note or running out of sound.

The problem? The musicians are unpredictable. The sun might hide behind a cloud, the wind might suddenly stop blowing, and the audience (the people using electricity) might suddenly start clapping louder in the evening.

This paper introduces a new, super-fast conductor named ELM (Extreme Learning Machine) who can predict exactly what this orchestra will play for the next few hours, helping the grid stay stable.

Here is the breakdown of how they did it, using simple analogies:

1. The Challenge: The "Island" Orchestra

The researchers studied the island of Corsica. Think of an island grid like a small, isolated band. They can't easily borrow instruments from the next town over (importing electricity is limited). If the wind stops, they can't just flip a switch to a giant nuclear plant; they have to rely on a mix of local sources.

• The Goal: Predict how much power will be made and used for the next 1 to 10 hours.
• The Old Way: The "Persistence" method. This is like guessing, "If it's sunny right now, it will be sunny in an hour." It's simple, but often wrong when the weather changes fast.
• The Complex Way: Using Deep Learning (like LSTM). This is like hiring a genius music theorist who studies every note ever played. It's powerful, but it takes days to train, requires a supercomputer, and even the theorist doesn't always know why they made a prediction.

2. The Solution: The "ELM" Conductor

The authors proposed using an Extreme Learning Machine (ELM).

• The Analogy: Imagine a chef who doesn't need to taste every dish to learn the recipe. Instead, they throw all the ingredients into a pot, stir them once with a specific, random motion, and poof—they instantly know the perfect seasoning.
• How it works: Unlike traditional AI that learns slowly by trial and error (like a student studying for years), ELM learns in a single step. It looks at the past data, makes a quick calculation, and gives an answer. It's incredibly fast and doesn't need a supercomputer.

3. The "MIMO" Trick: Seeing the Whole Picture

Most forecasters look at one instrument at a time (e.g., "How much wind will there be?"). This paper used a MIMO (Multiple-Input, Multiple-Output) approach.

• The Analogy: Instead of asking the drummer, the guitarist, and the singer separately what they will do, the ELM conductor asks the whole band at once.
• Why it helps: If the wind stops (bad for the wind turbine), the thermal generator might kick in to fill the gap. By looking at all the instruments together, the model understands that "If Wind goes down, Thermal goes up." This "teamwork" view makes the prediction more accurate than looking at each source in isolation.

4. The Results: Fast, Accurate, and Honest

The researchers tested this new conductor against the old methods using 6 years of real data from Corsica.

• Speed: The ELM trained in 104 seconds. The complex Deep Learning (LSTM) model took 44 minutes (about 25 times longer) on the same computer.
• Accuracy: For the next 5 hours, ELM was a star performer.
  • Solar & Thermal: It predicted these with incredible precision (over 98% accuracy).
  • Wind: This was the tricky instrument. The wind is chaotic, so the prediction was less perfect, but still better than just guessing.
  • Total Power: The model was excellent at predicting the total energy needed, which is the most important number for keeping the lights on.
• The "Magic" of Errors: Interestingly, when the model guessed wrong on solar power, it often guessed wrong on thermal power in the opposite direction. These errors canceled each other out, making the total prediction very accurate. It's like a team where one person trips left, and another trips right, but the group stays balanced.

5. Why This Matters for You

You might wonder, "Why do I care about a computer model in Corsica?"

• Cheaper Bills: Better predictions mean less wasted energy and less need to burn expensive diesel generators as a backup.
• Greener Planet: If we can predict renewable energy (wind/sun) better, we can use more of it and less fossil fuel.
• Real-Time Decisions: Because ELM is so fast, it can be used in real-time. If a cloud passes over a solar farm, the system can instantly adjust the grid before the lights even flicker.

The Bottom Line

This paper proves you don't always need a "super-complex" AI to solve energy problems. Sometimes, a simple, fast, and clever approach (ELM) that looks at the whole picture (MIMO) works better, faster, and cheaper than the complicated giants.

It's like realizing that to predict the weather, you don't always need a satellite; sometimes, a smart local who knows how the wind and clouds interact is enough to get the job done.

Technical Summary

1. Problem Statement

The integration of renewable energy (RE) sources (solar, wind, hydro) into power grids introduces significant variability and non-stationarity, challenging grid stability and supply-demand balancing. Traditional statistical methods often fail to capture the complex, non-linear interactions between multiple energy sources and consumption patterns. Furthermore, while deep learning models (e.g., LSTM) offer high accuracy, they suffer from:

• High computational costs and long training times.
• Lack of interpretability (black-box nature).
• Difficulty in real-time deployment and online learning.

There is a specific need for a forecasting framework that can simultaneously predict multiple energy variables (Multi-Input Multi-Output or MIMO) with high accuracy, low computational overhead, and the ability to adapt to non-stationary data without extensive preprocessing.

2. Methodology

The authors propose a Multi-Input Multi-Output (MIMO) Extreme Learning Machine (ELM) framework for short-term energy forecasting.

A. Data Source and Context

• Location: Corsica, France (an isolated island grid with high renewable intermittency and limited interconnection).
• Dataset: Six years of hourly data (2018–2023) covering 368,256 data points.
• Variables: Solar, Wind, Hydro, Thermal, Bioenergy, Imported Electricity, and Total Consumption (Total Production).
• Preprocessing:
  • Sliding Window: A window of 48 hours (2 days) is used to capture temporal dependencies.
  • Feature Engineering: Cyclic time encoding using sine and cosine functions to represent the 24-hour cycle, preserving periodicity.
  • Gap Filling: Linear interpolation for missing data.
  • Dependency Analysis: Mutual Information (MI) was used to quantify non-linear dependencies between variables, confirming strong correlations (e.g., Thermal-Bioenergy, Hydro-Thermal).

B. The ELM Model Architecture

• Type: Feedforward neural network with a single hidden layer.
• Key Mechanism: Unlike traditional backpropagation, ELM randomly initializes input weights and biases. Output weights are calculated analytically using the Moore-Penrose pseudo-inverse ($W = H^+Y$), eliminating iterative optimization.
• Configuration:
  • Inputs: 338 features (7 energy variables $\times$ 48 time steps + 2 time indices).
  • Hidden Layer: 4096 neurons with ReLU activation.
  • Outputs: 7 predicted values (one for each energy variable) at the forecast horizon.
• Training Strategy: Multiple random initializations (50 runs) are performed to select the configuration with the lowest in-sample RMSE, ensuring robustness.

C. Comparative Frameworks

The study compares the proposed MIMO-ELM against:

1. Persistence Model: A naive baseline assuming future values equal the last observed value.
2. LSTM (Long Short-Term Memory): A deep learning benchmark.
3. SISO (Single-Input Single-Output): Training separate ELM models for each variable independently.

3. Key Contributions

1. Novel MIMO-ELM Framework: First application of a unified MIMO-ELM approach to simultaneously forecast diverse energy sources (renewable and non-renewable) and total consumption in a single model.
2. Handling Non-Stationarity: The model adapts to seasonal and dynamic trends via cyclic time encoding and sliding windows without requiring data stationarization.
3. Efficiency vs. Accuracy Trade-off: Demonstrates that ELM achieves competitive or superior accuracy to LSTM while reducing training time by orders of magnitude (approx. 25x faster in the study) and offering a closed-form solution.
4. Decision Support: The model is designed for real-time applications, including online learning, making it suitable for grid operators and microgrid management.

4. Results and Performance

The models were evaluated using nRMSE (normalized Root Mean Square Error), nMAE, R², and a Gain (G) metric compared to the persistence model.

• Overall Accuracy:

  • The ELM model achieved R² > 0.98 for Total Consumption and Thermal energy at a 1-hour horizon.
  • nRMSE for Total Consumption was exceptionally low (< 0.05 even at a 10-hour horizon).
  • Thermal Energy: nRMSE of 5.1% at 1-hour horizon.
  • Solar Energy: nRMSE of 17.9% at 1-hour horizon (challenging due to intermittency).
  • Wind Energy: Remained the most difficult to predict (nRMSE ~42% at 1h), though the model still outperformed persistence.

• Forecast Horizon:

  • High accuracy is maintained up to 5 hours. Beyond this, error increases, particularly for volatile renewable sources.
  • The optimal forecasting horizon for total energy gain was found to be 7 hours (Gain $\approx$ 0.81).

• Comparison with Persistence:

  • ELM showed significant gains (G > 0) for Total, Thermal, Solar, Hydro, and Imported energy.
  • Bioenergy showed negative gains, indicating that for low-capacity, stable sources, the persistence model is currently more efficient.

• ELM vs. LSTM:

  • Training Time: ELM took ~104 seconds; LSTM took ~2651 seconds (25x slower).
  • Performance: In this specific setup, ELM outperformed LSTM across all metrics (e.g., LSTM nRMSE for Total was 0.17 vs. ELM's 0.046 at 5h horizon). LSTM struggled with the limited dataset size and hyperparameter tuning.

• MIMO vs. SISO:

  • MIMO provided marginal but consistent improvements over SISO (up to ~4% error reduction for Hydro and Solar).
  • MIMO is computationally more efficient (104s vs. 240s for SISO) as it trains a single model rather than seven separate ones.
  • MIMO failed to improve Bioenergy forecasting, suggesting low cross-variable dependency for that specific source.

5. Significance and Implications

• Operational Efficiency: The proposed method offers a lightweight, fast, and accurate tool for grid operators to optimize dispatch, manage storage, and integrate renewables, particularly in isolated or microgrid systems.
• Scalability: The closed-form solution of ELM allows for easy deployment in real-time systems and online learning scenarios where data streams continuously.
• Cost-Effectiveness: By avoiding the computational overhead of deep learning, the approach makes advanced forecasting accessible without requiring high-performance computing clusters.
• Future Directions: The authors suggest integrating operational constraints (maintenance, reserves) and exploring hybrid models (ELM + Deep Learning) to further handle highly volatile sources and extend the forecasting horizon.

Conclusion: The study validates that a MIMO-based Extreme Learning Machine is a superior alternative to both traditional persistence models and complex deep learning architectures for short-term multi-source energy forecasting, offering an optimal balance of speed, accuracy, and interpretability.