Data-driven charging strategies for grid-beneficial, customer-oriented and battery-preserving electric mobility

Data-driven cha rging strategies fo r grid-b eneficial, customer-o riented and battery-p reserving electric mobilit y Ka rl Schw enk 1,2 , Tim Ha rr 2 , René Großmann 2 , Ricca rdo Remo Appino 1 , V eit Hagenmey er 1 , Ralf Mikut 1 1 Institute fo r Automation and Applied Info rmatics, Ka rlsruhe Institute of T echnology Ka rlsruhe, Germany E-Mail: ka rl.schwenk@kit.edu 2 eDrive Innovations, Daimler A G Sindelfingen, Germany E-Mail: ka rl.schwenk@daimler.com 1 Intro duction Electric V ehicle (EV) p enetration, renewable energies, and customer orientation of car manufacturers [ 1 ] enables synergies b et ween energy supply , v ehicle users, and the mobility sector. Ho wev er, also new challenges arise [ 2 ] which w e target to examine with three types of agen ts and their persp ectiv es: EV user, p ow er supplier and car man ufacturer. Although man y researc h pap ers in the literature describe single p erspectives, a threefold contemplation of their connections as shown in Fig. 1, is still missing. In the following, w e briefly presen t the single p ersp ectiv es and resp ectiv e in teractions. Eletric v ehicle user P o w er supplier Car man ufacturer requires/pa ys for electric energy c harging infrastructure/prices exp ects reliable pro ducts offers EV + charging services aggregated c harging demand comp ensation pa ymen t Figure 1: In teraction triangle of electric vehicle user, p ow er supplier, and car manufacturer. T o satisfy their individual needs, v ehicle users exp ect their EV s to b e just as reliable and conv enient to use, as kno wn from combustion engine cars [ 3 , 4 ]. Ho wev er, limited range and longer charging times of EV s complicate individual mobilit y , manifesting e.g. in r ange anxiety [ 2 , 5 ]. F or EV users also the p o wer supply in teraction changes, as not only home appliances but also mobility requires electric energy . Hence, a con venien t and cost-efficien t charging pro cess is hard to ac hiev e without automation technology , when using an EV on a daily basis [6]. P ow er suppliers (distribution grid op erators, energy retailers) target an efficient and fail-proof grid op eration, for whic h a reliable forecast and con trol of electric loads is desired [ 7 , 8 ]. How ev er, the intermitten t nature of renewable Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 1 energy pro duction endangers the utility grid through v oltage and frequency fluctuation [ 9 , 10 ]. The energy demand caused b y charging EV s might further amplify this effect [ 11 ]. T o a void these issues, p o wer suppliers may influence the EV users’ affection to connect their vehicles via dynamic energy prices [ 9 , 10 , 12 ]. A t the same time, m ultiple EV s connected to the grid—esp ecially if bi-directional charging is enabled [ 13 ]—help p o wer suppliers to level out im balances due to renewable energy generation [14]. Electric mobilit y also p oses new issues for car man ufacturers : During c harging and discharging of EV batteries a degradation ( b attery aging ) occurs [ 15 ], that correlates with a v alue depreciation of the en tire EV. Ho wev er, EV users’ satisfaction requires reliable and v alue-stable products, which car manufacturers aim to achiev e by offering services, such as c harging assistants [ 6 ]. The provided c harging strategies target simplified and sustainable EV usage b y considering individual customer needs and battery aging. The remainder of this pap er is structured as follows: T o identify and develop missing mo dels of the outlined problem, a general approac h is presented in Section 2. Then, Section 3 describ es an online learning framew ork for data acquisition, storage and application. In Section 4, w e prop ose tw o data-driven consumption mo dels. Finally , Section 5 giv es a brief summary and outlo ok on future w ork. 2 App roach Despite existing mo del predictiv e approac hes [ 16 , 17 ] that supp ort EV users while c harging, individualized c harging strategies are in general still inadequately dealt with in the literature. W e target to identify missing mo dels that quan tify p erspectives and interactions, for which Fig. 2 shows a schematic map (the green b o xes represen t unexplored mo dels requiring further ev aluation). Automated c hargin g algorithm Charging b eha vior mo del Consumption mo del Battery degradation mo del Dynamic energy c harging map Optimization CH EC PD AED Electric V ehicle Electric v ehicle user Car man ufacturer P o w er supplier b eha vioral data c harging plan EC driving information driv e s/ c harges c harging/driving information PD c harging information AED lo cal condition CH: Charing Habits EC: Estimated Consumption PD: Pr e dicted Degradation AED: Aggreg a t ed Energy Demand Figure 2: Schematic map of mo dels to describ e p erspectives and interactions, Automated charging algorithm (orange b o x) comprises missing mo dels (green b o xes) 2 Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 T o automatically create charging plans that satisfy the interests of pow er suppliers, EV users, and car manufacturers at the same time (cf. Section 1), an A utomated c harging algorithm will b e required (Fig. 2, orange b o x). A Charging b ehavior mo del (cf. Section 2.1) could determine Charging Habits (CH, Fig. 2, green arrow) based on b eha vioral data of Electric vehicle users . Thereby , charging plans provided to the EV user can b e adapted to individual requiremen ts. In order to obtain user- and route-sp ecific Estimated Consumption (EC, Fig. 2, blue arro ws) for the EV, a data-driv en Consumption mo del (cf. Section 4) may use driving information of the EV. The EC can also help to estimate the remaining driving range of the EV. F urther, charging plans should b e adapted to b e battery preserv ativ e b y means of a Battery degradation mo del (cf. Section 2.2) that provides a Predicted Degradation (PD, Fig. 2, red arrows) for sp ecific charging strategies. Using the PD, car manufacturers may also monitor the condition of operating EV batteries. A Dynamic charging energy map (cf. Section 2.3) could calculate the Aggregated Energy Demand (AED, Fig. 2, pink arro ws) based on charging information of several EV s. The P o wer supplier ma y then make use of the AED for impro ved load forecast precision. T ogether with information ab out momen tary grid load conditions obtained from Po wer suppliers, peak sha ving and load balancing applications can b e included in the charging plan. An Optimization approac h may finally conclude all the information and calculate a c harging plan for the EV user. 2.1 Cha rging b ehavior mo del The effectiveness of charging strategies dep ends on user acceptance, which in turn requires user-sp ecific charging strategies [ 6 ]. Therefore, a suitable c haracterization of users’ driving and c harging b eha vior is necessary [ 18 ]. Existing approac hes in the literature distinguish b et ween qualitativ e and quantitativ e ones. Psyc hological approac hes, e.g. deducted from customer surveys [ 19 , 2 , 20 ] allow to c haracterize basic user types. Therein, gamific ation and inc entivation [ 2 ] methods are developed to increase user acceptance. Quan titative approaches aim at predicting driving and c harging b eha vior of EV users, e.g. to infer energy demand [ 21 , 22 , 23 , 24 ]. Ho wev er, an adequate user in tegration, i.e. individualized charging strategies based on EV users’ c harging habits is not represented. T o attain a more universal c haracterization of the EV users’ charging behavior, different user t yp es need to be determined b y analyzing b ehavioral data of the EV user. Thereby , differen t user clusters can b e iden tified, for which individual charging habits can b e determined (cf. Fig. 2, left). T o clarify the EV users’ ob jectiv es, w e plan a customer surv ey in future work. Therein, the basic requiremen ts tow ards automated c harging strategies are inquired from i) customers using a conv en tional car, ii) customers ab out to acquire an EV, iii) customers just recently started using an EV, and iv) customers already using an EV for a longer time. Sp ecific questions on the users’ mobility requiremen ts, c harging habits, doubts, and exp ectations are supp osed to rev eal further requirements for individualized charging strategies. Subsequen tly , a data-driv en analysis of c harging pro cesses and EV user information may supp ort these findings and ma y allow creating a mathematical model. 2.2 Battery degradation mo del The in terest of car manufacturers focuses on the depreciation pro cess of the EV and its k ey factors. As the battery considerably contributes to the vehicle v alue [ 25 ], we particularly consider battery degradation. Present-da y mobility concepts (e.g. car sharing, leasing) comprise the vehicle battery to remain property of the man ufacturer [ 26 ]. V alue- stable batteries are thus ev en more relev an t to op erate economically . T o quantify battery aging, usually the State of Health 1 (SOH) is used [ 27 , 28 ]. Hitherto SOH mo dels ha ve limited practicability due to a complex execution [ 29 ] or inaccurate State of Charge (SOC) estimations [ 30 ]. Data-driv en metho ds, e.g. Bay esian net wor ks and neural netw orks [ 31 , 32 ] hold feasible alternativ es for SOH estimation [ 33 ]. Ho wev er, the existing approaches barely use user-related data, for whic h reason the influence of charging behavior on battery aging is not represen ted prop erly . Hence, we target to ev aluate the influence of charging strategies on battery degradation (cf. Fig. 2, center righ t). By means of a neural net work regression mo del, w e target to estimate the energy consumption for driving (cf. Section 4), not taking any battery aging influences into accoun t. The mo del trained on data from batteries without degradation can b e used to estimate the energy consumption for EV s with aged batteries. A discrepancy b etw een the estimation and the real consumption indicates a battery aging caused by increased internal losses. T ogether with a feature-augmen ted map, this mo del can additionally estimate energy consumption for c harging scheduling, that also considers individual driving styles and environmen tal conditions. Once this battery aging estimation is mature, also a 1 Ratio of actual usable capacit y in relation to the nominal battery capacity Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 3 dep endency on c harging influences could b e inferred. A concept v alidation has to b e pro ceeded with both sim ulation and real data. 2.3 Dynamic charging energy map T o matc h energy usage and generation, p o wer suppliers require information on the energy demand at a certain time and place, targeting a fault-free deliv ery of electric energy . This includes the demand due to charging EV s. Researc h pap ers to mo del the energy demand caused by c harging EV s already exist. They differentiate among approac hes to predict and control the grid condition [ 11 , 34 , 35 , 14 ], and approac hes for dimensioning and lo cating new c harging stations [ 36 , 37 ]. Stochastic assumptions to describe the EV user behavior are a ma jor deficiency of these approac hes. By combining information about op erating EV s, the b eha vior of their drivers, and the asso ciated charging requiremen ts, w e target to aggregate the energy demand caused b y charging vehicles and predict vehicles’ anticipated c harging lo cation. 2 EV s requiring c harging can b e consolidated to larger energy demands c haracterized by a load profile ov er time and lo cation. Then, an “energy map” can b e created, con taining the aggregated energy demand in real time, or a prediction for future p oints in time (cf. Fig. 2, righ t). The required energy can either b e acquired directly from the energy mark et, or the information can be passed to p o w er suppliers to handle the predicted load accordingly . F or concept v alidation, a sim ulation environmen t has to be developed. 3 Data framew o rk In the following, we propose a cloud-based framew ork to acquire, store and analyze vehicle data, as sho wn in Fig. 3. A telematic system records selected Controller Area Netw ork (CAN) signals for a fleet of connected EV s during driving and c harging. Applications, e.g. distributed functions structured in microservices require a reliable internet connection to pro vide historical and liv e data. Existing devices such as data loggers store data in the car lo cally . Ho w ever, data can only b e transmitted infrequently via infrastructure-bound connections (e.g. WLAN or LAN). A real-time vehicle status can thus not b e established and the large amount of data causes the data transmission to b e time-intensiv e. F or this reason, we use a single-b oard computer (SBC) equipped with a GPS antenna and mobile internet connection as a gatewa y b et ween v ehicle and bac k-end infrastructure. The SBC initially deco des and filters data with libraries stored on its internal memory . Via softw are packages, the SBC can b e up dated and controlled remotely not requiring a ph ysical contact to the SBC. The data transmission utilizes a Message Queuing T elemetry T ransp ort (MQTT) proto col follo wing a publish and subscrib e mec hanism 3 . A central MQTT brok er (Fig. 3, left) organizes the data distribution. This mechanism also enables communication in to the vehicle. Curren tly ten EV s provide 370 signals of in terest, which w e record with a 10-Hz sampling rate 4 . W e pro cess and store all collected data according to its structure (Fig. 3, middle). A relational database stores v ehicle data and meta-information. F urther, a do cument-orien ted database allows to store semi-structured signal time series data efficien tly . Additionally , a library database allows transforming v ehicle sp ecific data in to a standardized format. The stored data can subsequently b e used to build mo dels or provide stand-alone microservices with information (Fig. 3, right). Each model or microservice can mak e use of sev eral data sources, or communicate among each other, resp ectiv ely . A separate application database contains configurations and meta-information necessary for the microservice op eration. The cloud structure allo ws contin uit y in v arious asp ects. Models can b e easily implemen ted, c hanged, monitored and deleted. Soft ware changes can b e applied and deplo yed contin uously . Through a constant a v ailability , the mo dels can be fed with new data each time it is pro duced. This allo ws a contin uous dev elopment of the model with a updated data stream. Ho wev er, a contin uous av ailability also constitutes a risk in terms of information securit y [ 39 ], that needs to be handled accordingly . The modular structure allows to create and train sev eral mo dels in parallel. An Application Database con tains all mo dels and their in ternal, learned parameters (Fig. 3, right). Via a human mac hine interface, the mo dels and their learning progress can b e monitored and analyzed. 2 The access to EV information sub jects to the present data security legislation [38]. 3 participants can publish information others can subscrib e to 4 equals data stream of appro x. 12 Megabyte p er hour and vehicle. 4 Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 Figure 3: Scheme of framework for data acquisition via a MQTT-broker, storage in several databases and application of data-driven models or microservices. This allows to sup ervise the data acquisition, storage, and application. F urthermore, databases can b e examined, microservices and mo dels can b e adapted, and errors can b e handled accordingly . The interface also pro vides feedbac k to the v ehicle fleet and vehicle users. 4 Data-driven consumption mo del Here, we prop ose a data-driven consumption model as an exemplary application of the framew ork presen ted in Section 3. Consumption mo dels, as used in charging sc heduling, help to estimate energy demands for routes the user will driv e [ 40 ]. Ho w ever, existing estimators mostly utilize physical models neglecting influences of the driver and the en vironment. Suc h consumption mo dels yield relative estimation errors b et ween 2.52 % and 8.3 % [ 41 , 42 ]. A higher mo del accuracy allows to compute a more reliable and thoroughly adapted charging strategy . Thus, w e aim to design a mo del estimating the total consumed energy p er driv en distance for a giv en driving b eha vior, en vironmental condition, and battery state. Note, that we use the dimensionless State of Charge (SOC) as a metric for the consumed energy . 4.1 Data selection T o abstract driving and environmen tal influences w e use time series data recorded during driving of the EV s. W e differen tiate the input signals in the three categories: driving behavior, battery state, and environmen tal condition Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 5 (as shown in T able 1). The ac c eler ation tor que , br ake tor que , and r e cup er ation tor que are measured for fron t and rear axle, i.e. t wo signals for those v alues exist. Thus, w e obtain a total signal num ber of 𝑁 signal = 15 . T able 1: Input signals group ed by driving behavior, battery state and environmental condition Driving behavior Battery state En vironmental condition acceleration torque total electric current geodetic altitude brake torque battery voltage ambien t temp erature recuperation torque battery temperature lateral acceleration vehicle speed electric current driving longitudinal acceleration With a v ariety of driving situations and en vironmental conditions as input, we target a univ ersal estimator that generalizes w ell on arbitrary input data. The samples are sections with duration 𝑡 sec = 6 min of EV trips representing a typical ride. All trips with insufficien t duration or faulty signal v alues are discarded b eforehand. 5 T o av oid estimation errors due to random noise, the originally 10-Hz-sampled signals are aggregated b efore passing them to the mo del. F or eac h signal and for all v alues within an aggregation time p eriod 𝑡 agg = 1 min the mean is calculated. With 15 signals m ultiplied b y 6 data p oin ts p er trip section we obtain the extracted features 𝑥 1 ,.., 90 . Note that w e c hose the parameters 𝑡 sec and 𝑡 agg according to preliminary test results. 6 F or eac h data sample representing a trip section of 𝑡 sec = 6 min w e calculate a lab el Γ = 𝑒 − 𝑒 𝑜 − 𝑜 . (1) Therein, 𝑒 is the SOC at the start of the trip section, 𝑒 the SOC at the end. Similarly , 𝑜 is the mileage (in kilometers) at the start of the trip section, 𝑜 at the end, resp ectiv ely . 4.2 Mo dels T o mo del the dependency b et ween the input features 𝑥 1 ,.., 90 and output lab el Γ , w e design t wo models. 4.2.1 Mo del A F or the first Mo del A w e use a linear regression mo del ^ Γ = W · x + 𝑏, (2) with the input features x = ( 𝑥 1 , 𝑥 2 , ..., 𝑥 90 ) ⊤ of the aggregated signal time series, the weigh t matrix W ∈ R 1 × 90 , and the bias 𝑏 ∈ R . The w eights W and the bias 𝑏 are c hosen to yield a minim um mean squared error (cf. Section 4.3) b et w een the training samples and the regression [43]. 4.2.2 Mo del B F or the second Mo del B we design a neural netw ork regression model. The input la yer comprise 90 nodes representing the input features 𝑥 1 ,.., 90 . F urther, w e use five hidden lay ers in a triangular shape, as shown in Fig. 4. Each of the hidden lay er no des is activ ated through a Rectified Linear Unit (ReLU). The output lay er of the neural netw ork consists of one no de represen ting the estimated consumption ^ Γ . F or training the mo del w e pro cess all trips into training samples (cf. Section 4.1). Then, the data is fed to the mo dels in batc hes of 32 samples ov er 100 ep ochs. The mo del is implemen ted in Python [ 44 ] using K er as [ 45 ]. W e use the A dam optimizer [46] with a learning rate 𝛼 = 0 . 001 . 5 In following ev aluations, these special cases need to b e contemplated separately . 6 In future work, a more detailed ev aluation on the selection of these parameters is required. 6 Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 Figure 4: Schematic arc hitecture Mo del B, green: input no des, hidden lay er no des in grey , output no de in orange, total num b er of no des below each lay er. 4.3 Evaluation The models are trained with data obtained from eigh t out of ten EV s. F or v alidation, the data obtained from the remaining t wo EV s is used, i.e. we compare the estimated consumption ^ Γ with the actual consumption Γ . T o quan tify the mo del performance, we calculate the metrics Mean Absolute Error (MAE) MAE = 1 𝑁 𝑁  𝑛 =1 | Γ 𝑛 − ^ Γ 𝑛 | , (3) Relativ e Mean Absolute Error (RMAE) RMAE = 1 𝑁 𝑁  𝑛 =1 | Γ 𝑛 − ^ Γ 𝑛 Γ 𝑛 | , (4) Mean Squared Error (MSE) MSE = 1 𝑁 𝑁  𝑛 =1 (Γ 𝑛 − ^ Γ 𝑛 ) 2 , (5) and Ro ot Mean Squared Error (RMSE) RMSE = ⎯ ⎸ ⎸ ⎷ 1 𝑁 𝑁  𝑛 =1 (Γ 𝑛 − ^ Γ 𝑛 ) 2 . (6) T able 2 rep orts the v alidation results according to the error metrics (3) - (6) . The results show, that the neural net w ork Mo del B outperforms the linear Mo del A in all four metrics. Using Model B, the estimation yields a mean absolute error of 0 . 00461 km − 1 , i.e. for a trip of 1 km driv en distance, the battery level change is estimated with an a verage discrepancy of 0 . 461 % SOC compared to the true v alue. On the contrary , Mo del A estimates the battery level c hange for suc h a trip with an av erage discrepancy of 52 . 783 % SOC. Note that an SOC of 100 % represen ts a fully c harged Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 7 T able 2: Evaluation of Mo del A and Mo del B with MAE, RMAE, MSE, and RMSE. Mo del MAE [km − 1 ] RMAE [-] MSE [km − 2 ] RMSE [km − 1 ] Mo del A 0.52783 1.45031 0.44548 0.66744 Mo del B 0.00461 0.01782 0.00004 0.00651 battery and an SOC of 0 % an empty battery , respectively . Considering this fact, the estimations obtained from Mo del A do not pro vide useful information on the EV’s consumption. Sho wing the results graphically further emphasizes this finding: Figure 5 illustrates the v alidation results of Mo del A. F or eac h v alidation sample with consumption Γ the estimation ^ Γ that the mo del pro vided is depicted as red dot. The green line represen ts an ideal mo del b eha vior for comparison. The green dashed lines represent an RMAE of 5%. It Figure 5: Estimated consumption ^ Γ (red dots) of Mo del A with v alidation data set compared with the true consumption Γ , ideal mo del output (green line), and 5% relative estimation error (green dashed lines). can b e seen, that most of the v alidation samples are not estimated correctly by the linear Mo del A, shown b y the red dots that are far from the green line. Based on the lo w accuracy of the linear mo del, we assume the real relation b et w een the input features 𝑥 1 ,.., 90 and the consumption Γ to b e non-linear. The results of the neural net work Model B supp ort this assumption: In the same manner as done for Mo del A, Fig. 6 sho ws the estimations for Mo del B. The results sho w a higher precision, as the estimations ^ Γ are lo cated closer to the ideal mo del represen tation, i.e. the green line. A mean relative estimation error 𝜖 B = 1 . 78 % (cf. T able 2) on the v alidation data supports this result. According to the results, the internal structure of the neural netw ork Mo del B seems to represent the influences on the energy consumption with muc h higher precision than the linear Mo del A. 8 Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 Figure 6: Estimated consumption ^ Γ (red dots) of Mo del B with v alidation data set compared with the true consumption Γ , ideal mo del output (green line), and 5% relative estimation error (green dashed lines). A further impro vemen t of the mo del performance is exp ected if further data, e.g. the actual w eigh t of the electric v ehicle due to baggage and passengers, would b e considered. Cross-v alidation o ver several vehicles should b e pro ceeded to test the generalization of the dev elop ed mo del. Note that the input data excludes signals allowing to infer the state of battery degradation. Thus, the consumption mo del does not consider effects of battery aging. Prosp ectiv ely , battery aging can b e estimated by training a similar mo del with data solely obtained from EV s, whose batteries hav e not degraded yet. Then, a consumption can b e estimated for vehicles that are assumed to hav e an aged battery . A discrepancy b etw een the estimation and the real consumption—b ey ond the kno wn estimation error—indicates a battery aging due to increased internal losses. This metho d would allow an en tirely data-driv en battery aging estimation based on data an ywa y created within the v ehicle op eration. Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 9 5 Conclusion and p ersp ective In this pap er, w e examined p erspectives and in teractions of Electric V ehicle (EV) users, car manufacturers and p o wer suppliers. W e proposed a concept to quan tify the objectives of all parties, aiming at automated calculation of EV c harging strategies. In this context, we outlined a cloud-based framew ork for data acquisition, storage and application as a common basis for data-driv en analyses. As an example use case, we proposed tw o data-driven mo dels based on linear regression and neural-net w ork regression to estimate sp ecific consumption, i.e. consumed energy p er driven distance. Therefore, we used time series data collected from a fleet of ten connected EV s. F or the mo del training only eigh t out of ten EV s were used, while the remaining t wo EV s were used for v alidation. The linear mo del seems to inadequately represent the true relation betw een input features and consumption. How ever, the neural netw ork mo del with five hidden la yers yields a mean relative absolute error of 1.78%. Considering the underlying data, the prop osed mo del outperforms estimators based on physical models as describ ed in the literature. More elaborate mo del v alidation with differen t data and other models, such as p olynomial mo dels, or other neural net work arc hitectures may allow further model precision improv ement. In future w ork, a mo del as proposed could estimate battery aging, if the training data is solely obtained from EV s, whose batteries hav e not degraded yet. Comparing the estimated consumption with the actual consumption for EV s, whic h ha ve degraded batteries, migh t indicate battery aging due to increased internal losses. This method w ould allo w an entirely data-driv en battery aging estimation. F urthermore, the remaining mo del parts of the prop osed concept require a more elab orate implemen tation and ev aluation. The in teractions among EV user, p o wer supplier, and car manufacturer should b e analyzed and describ ed mathematically . Finally , the findings need to b e com bined in an optimization approach, enabling the automated calculation of individualized EV charging strategies considering momentary grid load and battery preserv ation. F or concept ev aluation, we plan to use an exp erimen tal EV fleet within measure campaigns in the Ener gy L ab 2.0 [ 47 ]. Concluding, the feasibilit y of grid-optimal, battery-preserving and individualized c harging strategies needs to b e in vestigated with adequate indexes, suc h as user acceptance. References [1] In ternational Energy Agency, “Global EV outlo ok 2018: T o wards cross-modal electrification,” 2018. [2] M. Eider, M. Stolba, D. Sellner, A. Berl, R. Basmadjian, H. de Meer, S. Klingert, T. Sc hulze, F. Kutzner, and C. Kacp erski, “Seamless electromobility ,” in Pr o c e e dings of the Eighth International Confer enc e on F utur e Ener gy Systems - e-Ener gy 17 . ACM Press, 2017. [3] G. Broadb en t, G. Metternic ht, and D. Drozdzewski, “An analysis of consumer incen tiv es in supp ort of electric v ehicle uptake: An A ustralian case study ,” W orld Ele ctric V ehicle Journal , vol. 10, no. 1, p. 11, 2019. [4] M. Lützenberger, N. Masuch, T. Küster, D. F reund, M. V oß, C.-E. Hrabia, D. Pozo, J. Fähndric h, F. T rollmann, J. Keiser, and S. Albayrak, “A common approach to intelligen t energy and mobility services in a smart city en vironment,” Journal of A mbient Intel ligenc e and Humanize d Computing , vol. 6, no. 3, pp. 337–350, 2015. [5] M. Cuchý, M. Štolba, and M. Jakob, “Whole da y mobility planning with electric vehicles,” in Pr o c e e dings of the 10th International Confer enc e on A gents and A rtificial Intel ligenc e . SCITEPRESS - Science and T echnology Publications, 2018. [6] K. Sch w enk, M. F aix, R. Mikut, V. Hagenmeyer, and R. R. Appino, “On calendar-based scheduling for user- friendly c harging of plug-in electric v ehicles,” in 2019 2nd IEEE Conne cte d and A utonomous V ehicles Symp osium (CA VS) , 2019. [7] L. Liu, F. Kong, X. Liu, Y. Peng, and Q. W ang, “A review on electric vehicles in teracting with renewable energy in smart grid,” R enewable and Sustainable Ener gy R eviews , v ol. 51, pp. 648–661, 2015. [8] K. M. T an, V. K. Ramac handaramurth y , and J. Y. Y ong, “In tegration of electric vehicles in smart grid: A review on vehicle to grid technologies and optimization tec hniques,” R enewable and Sustainable Ener gy R eviews , vol. 53, pp. 720–732, 2016. 10 Pro c. 29. W orkshop Computational Intelligence, Dortmund, 28.-29.11.2019 [9] M. Allison, E. Akakab ota, and G. Pillai, “F uture load profiles under scenarios of increasing renew able generation and electric transp ort,” in 2018 5th International Confer enc e on R enewable Ener gy: Gener ation and Applic ations (ICREGA) . IEEE, 2018. [10] P . Aliasghari, B. Mohammadi-Iv atlo o, M. Alipour, M. Abap our, and K. Zare, “Optimal sc heduling of plug-in electric vehicles and renew able micro-grid in energy and reserve mark ets considering demand resp onse program,” Journal of Cle aner Pr o duction , v ol. 186, pp. 293–303, 2018. [11] R. Jarvis and P . Moses, “Smart grid congestion caused by plug-in electric vehicle charging,” in 2019 IEEE T exas Power and Ener gy Confer enc e (TPEC) . IEEE, 2019. [12] M. Alip our, B. Mohammadi-Iv atlo o, M. Moradi-Dalv and, and K. Zare, “Sto c hastic scheduling of aggregators of plug-in electric v ehicles for participation in energy and ancillary service markets,” Ener gy , vol. 118, pp. 1168–1179, 2017. [13] M. R. Mozafar, M. H. Moradi, and M. H. Amini, “A simultaneous approac h for optimal allo cation of renewable energy sources and electric vehicle charging stations in smart grids based on impro ved GA-PSO algorithm,” Sustainable Cities and So ciety , vol. 32, pp. 627–637, 2017. [14] R. R. Appino, M. Munoz-Ortiz, J. A. G. Ordiano, R. Mikut, V. Hagenmey er, and T. F aulw asser, “Reliable dispatc h of renewable generation via charging of time-v arying PEV p opulations,” IEEE T r ansactions on Power Systems , v ol. 34, no. 2, pp. 1558–1568, 2019. [15] L. K. Maia, L. Drünert, F. L. Man tia, and E. Zonderv an, “Expanding the lifetime of Li-ion batteries through optimization of c harging profiles,” Journal of Cle aner Pr o duction , v ol. 225, pp. 928–938, 2019. [16] A. D. Giorgio, F. Lib erati, and S. Canale, “Electric vehicles charging con trol in a smart grid: A model predictiv e con trol approach,” Contr ol Engine ering Pr actic e , v ol. 22, pp. 147–162, 2014. [17] A. Rajaei, M. Jo oshaki, M. F otuhi-Firuzabad, and M. Mo eini-Agh taie, “Enhancing p o wer distribution system flexibilit y using electric v ehicle charging management,” in 2019 27th Ir anian Confer enc e on Ele ctric al Engine ering (ICEE) . IEEE, 2019. [18] N. Kühl, M. Goutier, A. Ensslen, and P . Jo c hem, “Literature vs. twitter: Empirical insigh ts on customer needs in e-mobilit y ,” Journal of Cle aner Pr o duction , vol. 213, pp. 508–520, 2019. [19] E. Azadfar, V. Sreeram, and D. Harries, “The inv estigation of the ma jor factors influencing plug-in electric vehicle driving patterns and c harging behaviour,” R enewable and Sustainable Energy R eviews , vol. 42, pp. 1065–1076, 2015. [20] D. Lüddeck e, C. Seidl, J. Schneider, and I. Sc haefer, “Mo deling con text-aw are and inten tion-a ware in-car infotainmen t systems,” Softwar e & Systems Mo deling , v ol. 17, no. 3, pp. 973–987, 2016. [21] R. R. Desai, R. B. Chen, and W. Armington, “A pattern analysis of daily electric vehicle charging profiles: Op erational efficiency and environmen tal impacts,” Journal of A dvanc e d T r ansp ortation , v ol. 2018, pp. 1–15, 2018. [22] X. Li, Q. Zhang, Z. Peng, A. W ang, and W. W ang, “A data-driv en tw o-lev el clustering mo del for driving pattern analysis of electric v ehicles and a case study ,” Journal of Cle aner Pr o duction , vol. 206, pp. 827–837, 2019. [23] W. Dong, J. Li, R. Y ao, C. Li, T. Y uan, and L. W ang, “Characterizing driving styles with deep learning,” IBM Researc h China, 2016. [24] S. Sun, J. Zhang, J. Bi, and Y. W ang, “A machine learning metho d for predicting driving range of battery electric v ehicles,” Journal of A dvanc e d T r ansp ortation , vol. 2019, pp. 1–14, 2019. [25] J. Neubauer, A. Bro ok er, and E. W o od, “Sensitivity of battery electric vehicle economics to drive patterns, v ehicle range, and charge strategies,” Journal of Power Sour c es , v ol. 209, pp. 269–277, 2012. [26] G. Smith, J. So chor, and I. M. Karlsson, “Mobility as a service: Dev elopment scenarios and implications for public transp ort,” R ese ar ch in T r ansp ortation Ec onomics , vol. 69, pp. 592–599, 2018. [27] C.-H. Lee and C.-H. W u, “A nov el big data modeling metho d for impro ving driving range estimation of EV s,” IEEE A c c ess , vol . 3, pp. 1980–1993, 2015. Proc. 29. W orkshop Computational Intelligence, Dortm und, 28.-29.11.2019 11 [28] Y. Zheng, J. W ang, C. Qin, L. Lu, X. Han, and M. Ouyang, “A nov el capacit y estimation metho d based on c harging curve sections for lithium-ion batteries in electric v ehicles,” Ener gy , v ol. 185, pp. 361–371, 2019. [29] A. Eddahech, O. Briat, N. Bertrand, J.-Y. Delétage, and J.-M. Vinassa, “Beha vior and state-of-health monitoring of Li-ion batteries using impedance spectroscopy and recurren t neural netw orks,” International Journal of Ele ctric al Power & Ener gy Systems , vol. 42, no. 1, pp. 487–494, 2012. [30] M. Hannan, M. Lipu, A. Hussain, and A. Mohamed, “A review of lithium-ion battery state of charge estimation and management system in electric vehicle applications: Challenges and recommendations,” R enewable and Sustainable Ener gy R eviews , vol. 78, pp. 834–854, 2017. [31] D. D. Susilo, A. Wido do, T. Prahasto, and M. Nizam, “State of health estimation of lithium-ion batteries based on com bination of gaussian distribution data and least squares supp ort v ector machines regression,” Materials Scienc e F orum , vol. 929, pp. 93–102, 2018. [32] Y. Li, S. Zhong, Q. Zhong, and K. Shi, “Lithium-ion battery state of health monitoring based on ensemble learning,” IEEE A c c ess , vol. 7, pp. 8754–8762, 2019. [33] M. Berecibar, I. Gandiaga, I. Villarreal, N. Omar, J. V. Mierlo, and P . V. den Bossc he, “Critical review of state of health estimation metho ds of Li-ion batteries for real applications,” R enewable and Sustainable Ener gy R eviews , v ol. 56, pp. 572–587, 2016. [34] O. Sundstrom and C. Binding, “Planning electric-driv e vehicle charging under constrained grid conditions,” in 2010 International Confer enc e on Power System T e chnolo gy . IEEE, 2010. [35] M. Ma jidpour, C. Qiu, P . Chu, H. R. P ota, and R. Gadh, “F orecasting the EV c harging load based on customer profile or station measuremen t?” Applie d Ener gy , v ol. 163, pp. 134–141, 2016. [36] J. González, R. Alv aro, C. Gamallo, M. F uen tes, J. F raile-Ardanuy , L. Knapen, and D. Janssens, “Determining Electric V ehicle Charging Poin t Lo cations Considering Drivers’ Daily Activities,” Pr o c e dia Computer Scienc e , v ol. 32, pp. 647–654, 2014. [37] J. Li, X. Sun, Q. Liu, W. Zheng, H. Liu, and J. A. Stanko vic, “Planning Electric V ehicle Charging Stations Based on User Charging Beha vior,” in 2018 IEEE/A CM Thir d International Confer enc e on Internet-of-Things Design and Implementation (IoTDI) . IEEE, 2018. [38] Eurpäisc hes Parlamen t und EU Rat, “Datensch utz-Grundv erordnung, VER ORDNUNG (EU) 2016/679,” 2018. [Online]. A v ailable: h ttps://dsgvo- gesetz.de/ [39] D. Chen and H. Zhao, “Data securit y and priv acy protection issues in cloud computing,” in 2012 International Confer enc e on Computer Scienc e and Ele ctr onics Engine ering . IEEE, 2012. [40] M. Baum, J. Dibbelt, A. Gemsa, D. W agner, and T. Zündorf, “Shortest feasible paths with c harging stops for battery electric v ehicles,” T r ansp ortation Scienc e , 2019. [41] C. Fiori, K. Ahn, and H. A. Rakha, “Po w er-based electric v ehicle energy consumption mo del: Model dev elopment and v alidation,” A pplie d Ener gy , vol. 168, pp. 257–268, 2016. [42] J. Hong, S. Park, and N. Chang, “Accurate remaining range estimation for electric v ehicles,” in 2016 21st A sia and South Pacific Design A utomation Confer enc e (ASP-DA C) . IEEE, 2016. [43] S. W eisb erg, A pplie d line ar r e gr ession , 3rd ed. Wiley-In terscience, 2005. [44] Python Soft ware F oundation, “Python,” 2019. [Online]. A v ailable: https://www.p ython.org/ [45] F. Chollet et al. , “Keras,” 2015. [Online]. A v ailable: h ttps://keras.io [46] D. Kingma and J. Ba, “A dam: A metho d for sto c hastic optimization,” 2014. [47] V. Hagenmey er, H. K. Cakmak, C. Düpmeier, T. F aulw asser, J. Isele, H. B. Keller, P . Kohlhepp, U. Kühnapfel, U. Stucky , S. W aczowicz, and R. Mikut, “Information and comm unication technology in Energy Lab 2.0: Smart energies system sim ulation and control cen ter with an Op en-Street-Map-based pow er flow simulation example,” Ener gy T e chnolo gy , vol. 4, pp. 145–162, 2016. 12 Pro c. 29. W orkshop Computational Intelligence, Dortmund, 28.-29.11.2019