component increases the accuracy of the simulations and planning . Our innovative approach demonstrates that smarter optimization of power markets has the potential to drastically reduce consumers ’ electricity bills .
II . FORECASTING MODELS
In this study , the employed methodology encompasses various phases , ranging from load and photovoltaic yield forecasting to the optimization of battery charge / discharge control . Here , several forecasting models have been utilized , and their details are provided below :
PV yield forecasting : For the optimal operation of residential PV-battery systems , accurate predictions of PV yield and load profiles are required . PV yield forecasting is accomplished by combining weather forecast APIs with Sandia- and NREL-based physical models from PVLIB [ 8 ]. The predicted AC power yield profiles are generated using PVLIB 0.9.5 for an existing residential installation located in L Flanders , Belgium . The PV installation is modeled using its physical properties , 13 panels of 400 WP oriented towards the East ( surface azimuth of 83 °), 13 panels of 400 WP oriented towards the South ( surface azimuth of 263 °) and a solar panel tilt of 20 °. The PV system is connected to the grid via a Huawei SUN2000-6KTL-M0 inverter . The local Global Horizontal Irradiance ( GHI ), wind speed and temperature are provided daily via the Solcast Application Programming Interface ( API ) by providing the latitude and longitude coordinates . The measured PV yield is obtained by logging the inverter ’ s output AC Power via the available RS485 interface on a 1s interval .
Electric load forecasting : The electric load profile is a more complex signal due to consumers ’ non-linear and stochastic behavior . The timing of individual peak demands and peak injections leads to congestion in the electrical grid . Although extensive research has been conducted in this field , there is still a need for effective forecasting models . Our proposed framework includes a comprehensive analysis of short-term load forecasting based on authentic load profiles of Flemish customers , as provided by the DSO . Additionally , the accuracy of peak load forecasting must be adequately represented by common evaluation metrics . We have developed a model comparison framework that enables us to incorporate new models and additional metrics for assessing a variety of time series models in load prediction .
To objectively compare our models with different kinds of metrics , we introduce a model comparison framework written in Python . This framework offers flexibility to add custom models and metrics and ensures all tests are performed equally by providing the same input data for each prediction . To further ensure correctness , data leakage from training data to test data will automatically be prevented . In this framework , we have implemented various models , which will be discussed further .
Multiple methods have been shown to accurately predict day-ahead load consumption . Several algorithms were studied to evaluate their performance on an authentic load profile dataset , ranging from deep learning models ( Recurrent Neural Networks ( LSTM [ 9 ], GRU [ 10 ]), Hybrid Neural Networks ( CNNs followed by LSTMs [ 11 ]), N-BEATS [ 12 ], and
Temporal Fusion Transformers [ 13 ]), to machine learning models ( e . g ., XGBoost [ 14 ]) and regularized regression methods ( Elastic Net [ 15 ], Ridge [ 16 ], Lasso [ 17 ]), as well as clustering models and statistical autoregressive models such as ARIMA ( Dynamic Harmonic Regression [ 18 ]), Holt-Winters [ 19 ], and General Additive Model ( GAM ) [ 20 ]. Two approaches were used for day-ahead forecasting :
( 1 ) recursive multi-step forecast , in which we predict one point at a time in the day ahead , and take the prediction as the input and the last available kWh-value for the next point of the same day ahead , or
( 2 ) multiple output strategy : predicting all 96 points at once . Predictions from the different models were evaluated based on the four selected metrics .
All models were compared against a naive baseline model where the previous day is used as a prediction for the following day . We restrict the results to the top 5 models that outperform the baseline model in one or more metrics .
Conventional evaluation metrics like Mean Squared Error ( MSE ) and Mean Absolute Error ( MAE ) are used . Mean Absolute Percentage Error ( MAPE ) was discarded , due to zero values in the dataset which rendered results useless . Two novel metrics were introduced to focus on the specific needs in this project : the Absolute Cumulative Daily Error ( ACDE ) and the Peak Weighted Mean Square Error ( PWMSE ).
ACDE uses the absolute difference between the daily sum of the observed load y and the predicted load ŷ respectively for N daily values :
"
ACDE ( y , y () = | - y !
!#$
" − - y / % !#$
| ( 1 )
PWMSE is a weighted version of MSE , givingthe highest weight to the error corresponding to the largest peak in N daily observed load values ( actuals ):
PWMSE ( y , y () = - w ! ( y ! − y /) & %
"
!#$ with : weights w ! = y ′ ! / max ( y ′ ! , ε ) , normalized actuals y ′ ! =
' ! ()!*(')
( 2 )
) - .() - .(')()!*('), 0 ) and constant ε = 10 ( 1 to avoid zero-division .
III . DATASET USED
Authentic load profiles for Flemish households ( in Belgium ) are provided by Fluvius , the DSO , through an online open-source dataset , which we will refer to as the Open Fluvius dataset [ 21 ]. This dataset contains electricity consumption profiles for household customers , which can be utilized to predict future electricity demand . The dataset comprises electricity consumption data ( kWh values ) spanning one year for households in various regions of Flanders . This data is collected directly from the digital electricity meters of individual households and is sampled every 15 minutes . Each data point includes the measurement time , the electrical load drawn from the grid , and the injection into the grid .
,