Abstract
Accurate and effective power system load forecasting is an important prerequisite for the safe and stable operation of the power grid and the normal production and operation of society. In recent years, convolutional neural networks (CNNs) have been widely used in time series prediction due to their parallel computing and other characteristics, but it is difficult for CNNs to capture the relationship of sequence context and meanwhile, it easily leads to information leakage. To avoid the drawbacks of CNNs, we adopt a temporal convolutional network (TCN), specially designed for time series. TCN combines causal convolution, dilated convolution, and residual connection, and fully considers the causal correlation between historical data and future data. Considering that the power load data has strong periodicity and is greatly influenced by seasons and holidays, we adopt the Prophet model to decompose the load data and fit the trend component, season component, and holiday component. We use TCN and Prophet to forecast the power load data respectively, and then use the least square method to fuse the two models, and make use of their respective advantages to improve the forecasting accuracy. Experiments show that the proposed TCN-Prophet model has a higher prediction accuracy than the classic ARIMA, RNN, LSTM, GRU, and some ensemble models, and can provide more effective decision-making references for power grid departments.
Introduction
Accurate power load forecasting is a cornerstone of modern energy management. It allows grid operators to maintain a delicate balance between electricity supply and demand, ensuring system stability while minimizing operational costs and energy waste. However, power load data is notoriously difficult to model. It is characterized by complex, non-linear patterns driven by human activity and is heavily influenced by cyclical factors such as daily routines, weekly cycles, and seasonal shifts, as well as sudden anomalies like holidays.
Existing forecasting methods often fall short. Traditional statistical models like ARIMA are effective for linear trends but struggle with the non-linear complexities of modern grids. Recurrent Neural Networks (RNNs), including LSTMs and GRUs, are powerful but computationally expensive due to their sequential nature and can suffer from training instabilities over long sequences. While standard Convolutional Neural Networks (CNNs) offer parallelization, they are prone to "information leakage"—the use of future data points to predict past values—if not specifically architected for temporal causality. This paper introduces a hybrid framework that combines the strengths of Temporal Convolutional Networks (TCN) and Meta's Prophet model to overcome these challenges.
Methodology
The proposed TCN-Prophet model employs a parallel hybrid architecture. The time series data is processed simultaneously by two distinct branches, whose outputs are then fused using a weighted approach.
Temporal Convolutional Networks (TCN)
TCNs are a specialized class of CNNs designed for sequence modeling. They address the limitations of standard CNNs by employing causal convolutions, which ensure that the prediction at any given time step only depends on current and past information.
As illustrated in the figure above, a standard convolution can inadvertently incorporate future data (x_{t+1}) into the current output (y_t). TCNs solve this by using zero-padding to shift the kernel's receptive field, ensuring the model remains strictly causal.
In addition to causality, TCNs utilize dilated convolutions to exponentially increase the receptive field without a proportional increase in the number of layers. This allows the model to capture long-range dependencies in the power load history efficiently. Residual connections are also integrated to maintain training stability in deeper architectures.
The Prophet Model
While TCNs excel at non-linear feature extraction, the Prophet model provides robust handling of periodic components. Prophet is an additive model that decomposes a time series into three primary components: trend, seasonality (daily, weekly, yearly), and holidays. By explicitly modeling these "calendar effects," Prophet handles the predictable, human-driven cycles of power consumption, allowing the neural network branch to focus on more complex residuals.
Model Fusion via Least Squares
To combine the predictions from the TCN and Prophet branches, the authors utilize the Least Squares Method. This optimization technique determines the optimal weights (w1 and w2) for each model's output to minimize the total prediction error. This fusion strategy ensures that the final forecast leverages the best of both worlds: the deep learning-based feature extraction of TCN and the statistical rigor of Prophet.
Experimental Evaluation
The model was tested using hourly power load data from the ERCOT (Electric Reliability Council of Texas) dataset, covering eight regions from 2004 to 2021. This massive dataset (over 154,000 points per region) provided a rigorous testing ground for the hybrid approach. The authors used a continuous split for training, validation, and testing to prevent any temporal information leakage during evaluation.
Key Results
The TCN-Prophet model consistently outperformed a wide range of baselines, including ARIMA, LSTM, GRU, and standalone TCN/Prophet models.
- Accuracy Gains: Across all eight regions, the hybrid model achieved the lowest error metrics (MAPE, MSE, and MAE). In the COAST region, the hybrid approach improved forecasting accuracy (MAPE) by 19% compared to Prophet alone and 51% compared to a standalone TCN.
- Robustness: Qualitative analysis showed that the TCN-Prophet model tracked real-world load data much more closely during volatile periods and peak shifts compared to recurrent models.
Conclusion
The TCN-Prophet hybrid model represents a significant step forward in power load forecasting. By integrating causal deep learning with seasonal decomposition, it provides a more accurate and computationally efficient tool for grid management. Future research will focus on integrating external variables like weather conditions and the potential of attention-based Transformer architectures to further refine long-term forecasting capabilities.
Summary updated