FederatedNILM: A Distributed and Privacy-preserving Framework for Non-intrusive Load Monitoring based on Federated Deep Learning

The parameters of global deep learning model $w_{G}$, global sharing batch $B_{G}$, along with parameters of model training including learning rate $\eta$, momentum $\rho$ and loss function $\mathcal{L}$ need to be initialised first. Then set the global and local communication round index to 1, initialise the moment estimate variable $v$ to 0, and begin the first round of training.

The local households will train their own local deep learning model based on the local data $L_{n}$ after receiving the broadcast parameters from the global deep learning model. As described in $HouseholdsUpdate()$, in each local epoch, the local deep learning models aim to find the best approximation $F_{s}$ (Equation (4)). The local training returns the updated parameters from all the local models, which are then uploaded to the global cloud server.

The global deep learning model receives the local updated parameters from the global sharing batch $B_{G}$, and adopts federated averaging (FedAvg) to the parameter sets [9]. Then, the global deep learning model will be updated based on the FedAvg results. After this, the updated global parameters will be broadcast to the local models of each house. After running all the global communication rounds between the local households and the central cloud server, a final global deep learning model will be generated.

The input of the encoder is the household aggregated load consumption of the target appliances over a 12.6 minutes interval (the length of the input window of consumption datasets is 126). The encoder increases the output features from a single aggregation value to 256, while paying the price of decreasing the time signal resolution by 10 times.

The temporal pooling module consists of four average pooling modules. The filters in this module are reduced from the whole size of the input signal to one-sixth of it, which is the same case with the stride. After going through a convolutional layer, the feature dimension of the input is reduced to a quarter of its original size, and the acquired feature maps were upsampled to increase their size to the size of the input time signals. Then the upsampled feature maps (shallow features) are concatenated with the original input signal (deep features) from the temporal pooling to get the final feature maps. The fusion of the deep and shallow features of the temporal pool could enable this block to get contextual information fed into the decoder.

The decoder receives the output from the temporal pooling block and passes it to a convolutional layer to recover the temporal resolution. Then the output is fed into the final convolutional layer to produce the final output.

$Softmax$ is utilized to classify the states of multiple targeted appliances. After this, the binary cross-entropy is chosen as the loss function, which is given by:

where $N$ is the number of the training samples, $S$ is the length of the output from $Softmax$ classifier, $s_{n,s}^{i}$ is the true state of $i$-th appliance, and $p(s_{n,s}^{i})$ denotes the predicted probability that $i$-th appliance is in the activation state. Also, stochastic gradient descent (SGD) is selected as the optimizer to facilitate the convergence of the binary cross entropy.

The FederatedNILM model is implemented on Pytorch and conducted on a Windows 10 platform using an NVIDA GeForce RTX 2080 Ti GPU with 64GB of RAM and CUDA v10.2.

In this paper, we consider the UK-DALE dataset [26], which is a public and commonly used dataset and recorded both the aggregated load (high frequency and low frequency) and the individual appliances (low frequency) consumption from 5 different households (see Table 1).

To demonstrate the effectiveness of the proposed FederatedNILM model and the utilized deep neural network architecture, we deploy a simplified experimental setup following [4] for comparison purpose. We consider common appliances possessed by most houses, i.e. fridge, dishwasher, and washing machine, which narrowed the dataset to houses 1, 2 and 5. The date range selected from the three houses is from 12/04/2013 to 01/07/2015 for house 1, 22/05/2013 to 03/10/2013 for house 2, and 29/06/2014 to 01/09/2014 for house 5.

Following the same data preprocessing procedure in [4], abnormal load consumption records were firstly filtered out by the max power threshold provided in Table 2 followed by a down-sampling of the aggregated load from 1s to 6s to align with the submetered data. The resampled data were normalized by subtracting the mean and then dividing a constant value 2000 W. Finally, the state series of each target appliance were derived from the activation-time thresholding as the input of the proposed FederatedNILM model where relevant thresholds are provided in Table 2.

Both the seen house case and the unseen house case are considered to verify the effectiveness of FederatedNILM.

For the seen house case, the split of the three houses datasets is listed in Table 3. The first 80% series from each household were selected as the training set, followed by a 10% for validation, and 10% for testing. In this case, the disaggregation ability of the model is evaluated when signatures of specific appliances are learned.

For the unseen house case, the chosen of three houses allows us to train using two houses and test on another. The unseen house case aims to verify the generalization ability of the model, and the generic signature characteristics of the same type of appliances need to be distinguished. As detailed in Table 4, we split two houses data into training and validation sets, and use the other unseen house as the test set. Then, we average results of all three different combination cases of the training and testing for comparative analysis.

The parameters used in the FederatedNILM model training are listed in Table 5. Different global rounds [2, 4, 6, 8, 10] are selected to conduct comparative experiments. We repeat the experiment five times for all the cases and report the average performance for each model.

Four evaluation metrics are used to assess the performance of the proposed framework.

By denoting true positive as TP, true negative as TN, false positive as FP, and false negative as FN, the evaluation metrics are defined as follows:

where precision represents the proportion of TPs to all the data sequences classified to the ON state. Recall denotes the ratio of TPs to all data sequences that are actually in the ON state. Accuracy reflects the ratio of all correctly identified samples to all the data sequences. $F_{1}$ is defined as a weight average representation for precision and recall within the range of $[0,1]$. An $F_{1}$ close to 1 indicates that the classification results for the target appliance are better.

In this section, we conduct comparative studies between the proposed FederatedNILM and Centralized-NILM for the seen house case. We run 10 global rounds with 10 local epochs for FederatedNILM training, and 100 epochs for Centralized-NILM training.

Figure 3 shows the disaggregation performance in the seen house case.

From the results, we can see that for each appliance, both Centralized-NILM and FederatedNILM achieve satisfactory results on the dishwasher and washing machine, and reasonable results on the fridge. It is worth pointing out that the fridge consumes relatively low power compared with other appliances and is likely to be learned with less evident signature during model training since its consumption can easily be omitted as unidentified load noise. This might explain why testing results of the fridge show relatively low scores compared with other appliances.

It should be highlighted that the FederatedNILM achieves very similar performance to Centralized-NILM on all appliances. Therefore, it is reasonable to infer that our proposed FederatedNILM framework works well in a distributed and privacy-preserving manner for the seen house case and can achieve a good trade-off between data privacy and data utility.

In this section, we apply our proposed models for the unseen house case to evaluate the generalization ability of FederatedNILM.

Figure 4 shows the disaggregation scores on a house not seen during training for the three target appliances.

Compared with the seen house case, both models produce satisfying results for fridge and dishwasher but perform less well for washing machine. One possible reason is that the washing machine, as a multi-state appliance, has more complex signature characteristics [27], which is likely to be more difficult to distinguish the generic signature characteristics.

We can also find that the FederatedNILM has very similar performance compared with Centralized-NILM on all appliances for the unseen house case, which aligns well with our conclusion on the proposed FederatedNILM framework in terms of data privacy and data utility in Section 5.2.1 for the seen house case.

In this section, different global rounds and their corresponding training costs and testing results are given to explore the relationship between training efficiency and model performance.

The epochs for the Centralized-NILM are set to 20, 40, 60, 80, and 100 respectively corresponding to the global rounds 2, 4, 6, 8, 10 with 10 local epochs of the FederatedNILM model. Tables 6 to 8 listed the testing results of both models using different global rounds in the unseen house case for fridge, dishwasher, washing machine respectively, and the best results of each model are marked in bold. Figure. 5 shows the training time of the two models with the increasing of epochs/global rounds.

The overall performance of the FederatedNILM model and the Centralized-NILM generally improves with the increase of the epochs.

With the increase of the global rounds, the time consumption cost also increases. The training time consumption of the FederatedNILM model is slightly higher but still within the acceptable range compared with the Centralized-NILM. Besides, given similar classification scores, the proposed FederatedNILM model has similar training time efficiency compared with the Centralized-NILM. For instance, the proposed model obtained an accuracy, precision, recall, $F_{1}$ of 0.98, 0.77, 0.82, 0.78 respectively for the dishwasher when global rounds reach 10 whereas similar scores (0.98, 0.76, 0.90, 0.79 for accuracy, precision, recall, $F_{1}$, respectively) were achieved in the Centralized-NILM with 100 epochs. Therefore, we can also reasonably infer that the training cost of FederatedNILM is within an acceptable range compared with the Centralized-NILM under similar accuracy requirements. The above analysis can further confirm the proposed model could achieve a satisfactory trade-off between the computational costs and the model performance.

In the previous subsections, we compare the proposed FederatedNILM with Centralized-NILM to examine its feasibility from the perspective of communication cost and model accuracy (i.e. distributed vs centralized). In the following, we compare our adopted deep learning architecture (used in both FederatedNILM and Centralized-NILM) with state-of-the-arts considered in [4] on model generalization performance for NILM task.

By considering the experiment environment setup in this paper and in [4] and to allow for a direct comparison, the appliances and corresponding house numbers listed in Table 9 are considered.

Table 10 gives the comparative results of FederatedNILM and Centralized-NILM with state-of-the-arts for dishwasher and washing machine respectively. The models considered for the comparative study are combinatorial optimization(CO), factorial hidden Markov models (Neural-NILM FHMM), and deep neural network (Autoencoder, Rectangles, Neural-NILM LSTM, centralized-NILM, and FederatedNILM).

In general, the deep neural network-based models for NILM have relatively better performance than other models. It is also worth pointing out that both the Centralized-NILM and FederatedNILM achieved better testing results on the unseen house 2 (see Tables 9 and 10) than the average testing results of all three cases in Table 4 for the washing machine. One possible reason is that house 1 provides the largest dataset, and when this house is considered as an unseen house as in Table 4, the training size of the model becomes much smaller, which could impact the model performance. Among all the methods, the Centralized-NILM model achieved the highest accuracy score for both appliances, which proves that the deep learning architecture we utilized in this paper could enhance the local training performance and thus improve the overall state inference accuracy in our framework. As already discussed in previous subsections, the above results also demonstrated that our proposed FederatedNILM with the same deep learning architecture as in Centralized-NILM could achieve promising generalization performance. Considering FederatedNILM works in a distributed and privacy-preserving manner, which is fundamentally different from the other considered methods (i.e. centralized based methods), our proposed framework could achieve a good trade-off between data privacy protection and data utility.